[llvm-commits] [llvm] r137562 - /llvm/trunk/lib/CodeGen/SelectionDAG/SelectionDAG.cpp

Eli Friedman eli.friedman at gmail.com
Sat Aug 13 23:07:30 PDT 2011


On Sat, Aug 13, 2011 at 11:02 PM, Eli Friedman <eli.friedman at gmail.com> wrote:
> 2011/8/13 Rotem, Nadav <nadav.rotem at intel.com>:
>> Eli,
>>
>> You have a good point. I didn't think about it. How do you suggest we fix it ?
>
> Maybe expand it the same way DAGTypeLegalizer::ExpandOp_BUILD_VECTOR
> does?  That isn't really guaranteed to work, but it'll probably work
> in practice.
>
> Longer-term, we should probably add a ConstantVector SDNode and get
> rid of BUILD_VECTOR; that would get rid of a bunch of weirdness in
> SelectionDAGs.

I'm just realizing from your comments in PR10653 that you're getting
into a situation that should be impossible to get into, though... it
shouldn't be possible to end up with a SIGN_EXTEND_INREG that isn't
legal after vector op legalization.

-Eli




More information about the llvm-commits mailing list