[Bug tree-optimization/49002] New: 128-bit AVX load incorrectly becomes 256-bit AVX load

[athena at fftw dot org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49002

           Summary: 128-bit AVX load incorrectly becomes 256-bit AVX load
           Product: gcc
           Version: 4.6.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: tree-optimization
        AssignedTo: unassigned@gcc.gnu.org
        ReportedBy: athena@fftw.org


Created attachment 24249
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=24249
test case for this bug

The attached foo.c program contains two 128-bit AVX loads from
locations of type __m128d.  However, gcc-4.6 converts the first
load into a 256-load.  While this is semantically correct, because
the upper 128-bits are ignored, the vmovapd instruction has different
alignment requirements in 128 and 256-bit mode, and therefore this
conversion causes spurious segfaults when the data is not 32-byte aligned.

Compile the attached program as follows:

   x86_64-linux-gnu-gcc-4.6 -mavx -O -S foo.c

The generated assembly contains the incorrect load:

        vmovapd (%rdi), %ymm0

By contrast, gcc-4.5 generates the correct 128-bit load instruction "vmovapd
(%rdi), %xmm0".