AMDGPU would like to have MVTs for v3i32, v3f32, v5i32, v5f32. This
commit does not add them, but makes preparatory changes:
- Exclude non-legal non-power-of-2 vector types from ComputeRegisterProp mechanism in TargetLoweringBase::getTypeConversion.
- Cope with SETCC and VSELECT for odd-width i1 vector when the other vectors are legal type.
- Fixed an assumption of power-of-2 vector type in ARM.
- Fixed assumptions of power-of-2 vector type in AMDGPU kernel arg handling.
- Fixed AMDGPU cost analysis to behave the same.
Some of this patch is from Matt Arsenault, also of AMD.