about auto-vectorization and run time stack align

[WANG.Jiong]

Hi all:

    Recently I found some tricky bugs in a private target which supports 
SIMD instructions.

    The reason is auto-vectorization will try to use those SIMD 
instructions which require data address be 32 bytes aligned.

    So a target specific option, which enable generating runtime stack 
check prologue, like x86's -mstackrealign is always needed if we want to 
make sure -ftree-vectorize work corrently?
    or are there something else I have missed?

    my gcc version is 4.3.5

    Thanks in advance.

---
Best,
KwongYuan