Patch for 4.7: Avoid subreg'ing VFP D registers in big-endian mode

[Ramana Radhakrishnan]

On Fri, Feb 15, 2013 at 10:40 PM, Seth LaForge <sethml@google.com> wrote:
> gcc 4.7.2 generates incorrect code for big-endian ARM VFP processors
> when storing a local double to a packed memory location, as described
> in bug 56351.
> http://gcc.gnu.org/bugzilla/show_bug.cgi?id=56351
>
> A fix has been submitted to the trunk by Julian Brown:
> http://gcc.gnu.org/git/?p=gcc.git;a=patch;h=4f0e4afcb2a9d345ed7e9c14233fb97d3a3e8d3c
>
> This fix is not included in the latest 4.7 snapshot.  Can we please
> include it in the 4.7 branch so that this serious codegen bug is fixed
> in 4.7.3?
>
> Unfortunately, it's a not a trivial backport, as it appears that
> support for ARM FPA floating point has been removed from gcc since
> 4.7.  Below is my best attempt at a backport - it works for me, on a
> big-endian VFP processor.  I'm not familiar with this code, though -
> Julian, can you take a look and confirm whether I'm doing this right?
>
> 2013-02-15 Seth LaForge <sethml@google.com>
>
>         * config/arm/arm.h (CANNOT_CHANGE_MODE_CLASS): Avoid subreg'ing
>         VFP D registers in big-endian mode.
>         Backported from a change by Julian Brown <julian@codesourcery.com>

This is not the correct form of a changelog entry . Please see other
examples of Backports in the Changelog file for the 4.7 branch.

Instead please use something like this.

<DATE>  <yourname>  <youremailid>

<TAB>Backport
<TAB>2012-10-22  Julian Brown  <julian@codesourcery.com>
	  * config/arm/arm.h (CANNOT_CHANGE_MODE_CLASS): Avoid subreg'ing
	  VFP D registers in big-endian mode.

Ok to backport provided no regressions  when running the testsuite in
big endian mode.

>
> diff -urp gcc-4.7.2/gcc/config/arm/arm.h gcc-4.7.2.new/gcc/config/arm/arm.h
> --- gcc-4.7.2/gcc/config/arm/arm.h      2012-01-09 20:14:09.000000000 -0800
> +++ gcc-4.7.2.new/gcc/config/arm/arm.h  2013-02-15 14:03:09.113531299 -0800
> @@ -1123,11 +1123,18 @@ enum reg_class
>  /* FPA registers can't do subreg as all values are reformatted to internal
>     precision.  In VFPv1, VFP registers could only be accessed in the mode
>     they were set, so subregs would be invalid there too.  However, we don't
> -   support VFPv1 at the moment, and the restriction was lifted in VFPv2.  */
> +   support VFPv1 at the moment, and the restriction was lifted in VFPv2.
> +   In big-endian mode, modes greater than word size (i.e. DFmode) are stored in
> +   VFP registers in little-endian order.  We can't describe that accurately to
> +   GCC, so avoid taking subregs of such values.  */
>  #define CANNOT_CHANGE_MODE_CLASS(FROM, TO, CLASS)              \
> -  (GET_MODE_SIZE (FROM) != GET_MODE_SIZE (TO)                  \
> -   ? reg_classes_intersect_p (FPA_REGS, (CLASS))               \
> -   : 0)
> +  (TARGET_VFP                                                  \
> +  ? TARGET_BIG_END                                             \
> +    && (GET_MODE_SIZE (FROM) > UNITS_PER_WORD                  \
> +       || GET_MODE_SIZE (TO) > UNITS_PER_WORD)                 \
> +    && reg_classes_intersect_p (VFP_REGS, (CLASS))             \
> +  : GET_MODE_SIZE (FROM) != GET_MODE_SIZE (TO)                 \
> +    && reg_classes_intersect_p (FPA_REGS, (CLASS)))

Watch out for alignment of the \'s though it might be your mailer
munging things rather than anything else. Can you please make sure
that the \'s are aligned to the same column please ? This is why I
prefer to get git to generate out patches using git-format-patch and
attach them as text files to all my mails otherwise it seems to mess
things up like this.

Thanks,
Ramana