Re: [arm-gnu] Possible bug in g++ frontend (G++ Lite arm-2009q1-203)

[Juha Kallioinen <juha.kallioinen@xxxxxxxxx>]

ext ext Daniel Jacobowitz wrote:
> On Tue, Oct 13, 2009 at 05:23:44PM +0300, Juha Kallioinen wrote:
> > -march=armv7-a -mcpu=cortex-a8 %{!mfpu:-mfpu=vfp} -mfloat-abi=softfp
>
> Does %{!mfpu=*:-mfpu=vfp} work better?

Yes, that works! Thanks :)

> > Similar line works with cc1 spec.
>
> There really should be no difference in specs processing between
> gcc/cc1 and g++/cc1plus.  I'm confused by your results; for example,
> the above shouldn't have worked in either case.

Well, I'm just reporting what I've observed. I'm using the downloaded 
toolchain as-is and in this case there is some difference in the way gcc and 
g++ follow the spec rules when the rules are identical.

To me it looks like g++ inserts the spec file rule twice and gcc only once 
and is this might be the reason why they behave differently. I don't know if 
this is the correct way to observe it, but it can be seen in the logs below.

Here's a snippet from the output of:

$ echo | bin/arm-none-linux-gnueabi-g++ -xc++ -specs=./gcc.specs -v -
...
gcc version 4.3.3 (Sourcery G++ Lite 2009q1-203)
COLLECT_GCC_OPTIONS='-v' '-shared-libgcc' '-funwind-tables'

/home/kaltsi/tmp/cs/arm-2009q1/bin/../libexec/gcc/arm-none-linux-gnueabi/4.3.3/cc1plus 
-quiet -v -iprefix 
/home/kaltsi/tmp/cs/arm-2009q1/bin/../lib/gcc/arm-none-linux-gnueabi/4.3.3/ 
-isysroot /home/kaltsi/tmp/cs/arm-2009q1/bin/../arm-none-linux-gnueabi/libc 
-D_GNU_SOURCE - -march=armv7-a -mcpu=cortex-a8 -mfpu=neon -mfloat-abi=softfp 
-fno-omit-frame-pointer -fno-optimize-sibling-calls 
-Wno-poison-system-directories -quiet -dumpbase - -auxbase - -version 
-funwind-tables -march=armv7-a -mcpu=cortex-a8 -mfpu=neon -mfloat-abi=softfp 
-fno-omit-frame-pointer -fno-optimize-sibling-calls 
-Wno-poison-system-directories -o /tmp/ccUSlt1v.s
...

See that the -march and -mcpu etc options that come from the specs file are 
there twice.

Now with gcc:

$  echo | bin/arm-none-linux-gnueabi-gcc -xc -specs=./gcc.specs -v -
...
gcc version 4.3.3 (Sourcery G++ Lite 2009q1-203)
COLLECT_GCC_OPTIONS='-v' '-funwind-tables'

/home/kaltsi/tmp/cs/arm-2009q1/bin/../libexec/gcc/arm-none-linux-gnueabi/4.3.3/cc1 
-quiet -v -iprefix 
/home/kaltsi/tmp/cs/arm-2009q1/bin/../lib/gcc/arm-none-linux-gnueabi/4.3.3/ 
-isysroot /home/kaltsi/tmp/cs/arm-2009q1/bin/../arm-none-linux-gnueabi/libc 
- -march=armv7-a -mcpu=cortex-a8 -mfpu=neon -mfloat-abi=softfp 
-fno-omit-frame-pointer -fno-optimize-sibling-calls 
-Wno-poison-system-directories -quiet -dumpbase - -auxbase - -version 
-funwind-tables -o /tmp/ccy7783x.s
...

those same options are there only once.

When I was using my original specs file, g++ would insert the -mfpu=vfp 
there twice and the latter one would override user's command line option. 
Gcc would not.

Here's what the same output looks like with the original spec file:

$ echo | bin/arm-none-linux-gnueabi-g++ -xc++ -specs=orig.specs -v -mfpu=neon -

...
gcc version 4.3.3 (Sourcery G++ Lite 2009q1-203)
COLLECT_GCC_OPTIONS='-v' '-mfpu=neon' '-shared-libgcc' '-funwind-tables'

/home/kaltsi/tmp/cs/arm-2009q1/bin/../libexec/gcc/arm-none-linux-gnueabi/4.3.3/cc1plus 
-quiet -v -iprefix 
/home/kaltsi/tmp/cs/arm-2009q1/bin/../lib/gcc/arm-none-linux-gnueabi/4.3.3/ 
-isysroot /home/kaltsi/tmp/cs/arm-2009q1/bin/../arm-none-linux-gnueabi/libc 
-D_GNU_SOURCE -isystem /usr/local/include -isystem /usr/include - 
-march=armv7-a -mcpu=cortex-a8 -mfloat-abi=softfp -fno-omit-frame-pointer 
-fno-optimize-sibling-calls -Wno-poison-system-directories -quiet -dumpbase 
- -mfpu=neon -auxbase - -version -funwind-tables -march=armv7-a 
-mcpu=cortex-a8 -mfpu=vfp -mfloat-abi=softfp -fno-omit-frame-pointer 
-fno-optimize-sibling-calls -Wno-poison-system-directories -o /tmp/ccbYJKSV.s
...

The -mfpu=vfp from the spec file ends up being the last mfpu option on the 
command line.

--
Juha Kallioinen