[llvm-commits] [llvm] r53035 - in /llvm/trunk: include/llvm/CodeGen/SelectionDAG.h lib/CodeGen/SelectionDAG/LegalizeDAG.cpp lib/CodeGen/SelectionDAG/SelectionDAG.cpp lib/CodeGen/SelectionDAG/SelectionDAGISel.cpp lib/Target/ARM/ARMISelLowering.cpp lib/Target/CellSPU/SPUISelLowering.cpp lib/Target/PowerPC/PPCISelLowering.cpp lib/Target/Sparc/SparcISelLowering.cpp lib/Target/X86/X86ISelLowering.cpp

[Mon P Wang]

Hi Duncan,

I ran into the problem when I was playing with "widening*" vector  
types.  It is something I haven't submitted a patch for so my test  
case will not work for you at this point.   What problem do you think  
will occur in LegalizeTypes? If you can let me know, I can do a quick  
look into it.

-- Mon Ping


* By widening a vector type, I mean assume one have a series of 3xi32  
vector operations when the hardware supports 4xi32 vector operations.   
In this case, it is more efficient to widen the 3xi32 to a 4xi32 (by  
copy the last element value in the 4th position) and using the 4xi32  
vector operations and undo the widening when one copies the results  
back to memory.


On Jul 4, 2008, at 11:41 AM, Duncan Sands wrote:

> Hi, do you have a testcase?  This problem might
> occur with LegalizeTypes too.
>
> Thanks,
>
> Duncan.