[llvm-commits] [PATCH] Fix assertion failure with shufflevectors on ARM

Stepan Dyatkovskiy stpworld at narod.ru
Wed Aug 15 02:06:26 PDT 2012


Hi, Eli
Eli Friedman wrote:
> This bit of math looks unsafe... can you actually guarantee this
> assertion holds?  It seems like it would be more straightforward to
> put the bitcast after the concat_vectors

Currently (without patch) LegalizeDAG got next node:

v16i8 concant_vectors v4i16, v4i16

It couses an assertion when it tryies to create extract_vector_element 
node, since indexing is wrong.
Patch adds operands bitcasting for concat_vectors, so the node will look 
like:
v16i8 concat_vectors v8i8, v8i8

Taking into account your remark, I'll try to fix output concat_vectors 
type to v8i16.
Is it what you mean?

-Stepan.



More information about the llvm-commits mailing list