Bitcast between 2 different SDNode vector types

Project Infrastructure LLVM Dev List Archives
1

Hello.
     I am trying to emulate in software in my SIMD processor back end vector types with elements i32 using the native vector types with i16 (for example: v64i32 with v128i16).
     For example, when I want to add 2 v64i32 I use first a getNode(ISD::BITCAST, ...) to typecast the operands of add. But this does not seem to work - I tried to get inspired from http://llvm.org/docs/doxygen/html/X86ISelLowering_8cpp_source.html. When I use that instruction I get an error saying "Cannot select [...] v128i16 = bitcast txy", where txy is of type t64i32.
   So I tried to declare at the beginning of the ISelLowering class:
     setOperationAction(ISD::BITCAST, MVT::v64i32, Expand /* or Promote */);
     AddPromotedToType (ISD::BITCAST, MVT::v64i32, MVT::v128i16); // and also MVT::v128i16, MVT::v64i32
   This eventually led to replacing during Instruction selection the
        v128i16 = bitcast t52
      with
        v128i16,ch = load<LD256[FixedStack3]> t110, FrameIndex:i64<3>, undef:i64
      but the load<LD256[FixedStack3]> operation is not really good for me.

     Can somebody help me? I need a simple bitcast operation that just changes the type of the vector SDNode without any further complication, if possible.

   Thank you,
     Alex

2

It seems to me that you’d actually want to keep the BITCAST’s legal and implement a pattern in the .td files for doing them.

For example, on PPC our vector bitcasts are no-ops (except that we prefer to keep some types of vectors in some of the register classes). You can see all the bitconvert patterns in lib/Target/PowerPC/PPCInstrVSX.td.

That will allow the instruction selector to select the bitcasts and won’t change things like the load you included in your message.