[Bug rtl-optimization/47246] New: [4.6 Regression] Invalid immediate offset for Thumb VFP store

["ibolton at gcc dot gnu.org" <gcc-bugzilla at gcc dot gnu dot org>]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=47246

           Summary: [4.6 Regression] Invalid immediate offset for Thumb
                    VFP store
           Product: gcc
           Version: 4.6.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: rtl-optimization
        AssignedTo: unassigned@gcc.gnu.org
        ReportedBy: ibolton@gcc.gnu.org
              Host: arm-linux-gnueabi
            Target: arm-linux-gnueabi
             Build: arm-linux-gnueabi


Created attachment 22939
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=22939
This is the preprocessed source that shows the issue

The following patch, which went into trunk as revision 168045, has made it
possible for gcc to generate invalid offsets for some VFP stores:

  http://gcc.gnu.org/ml/gcc-patches/2010-12/msg01349.html

The range -1024 to +1024 is ok for 's' co-processor registers, but not for
generic 'r' registers.

The attached preprocessed source can be made to show the issue by compiling as
follows:

  gcc -O3 -mthumb mcpu=cortex-a9 -mfloat-abi=softfp -mfpu=vfpv3-d16 besttry.i
--save-temps

The error caused is this:

  /tmp/ccxviNhV.s: Assembler messages:
  /tmp/ccxviNhV.s:37: Error: offset out of range

I have attached the produced besttry.s.

In the short-term, I suggest that we limit the negative offset, to be able to
cover the case where VFP store ends up using general registers.  I do not know
how we could allow a larger negative offset for the 's' registers, when it is
not known until later what type of register will get allocated.