[LLVMdev] question on scalarization

[Bjorn De Sutter]

Hi all,

I have a backend for an in house architecture, and would like to start working on support for vector datatypes and intrinsics. As I starter I would like to have vector code scalarized, as is done, e.g., for the Mips backend. However, I cannot find a way to force the instruction selector or type legalizer to scalarize the vectors (i.e., vector types and vector ops). Can anyone help me out?

The code I am trying to compile is as follows:

typedef short v2sw __attribute__ ((vector_size (8)));

int main(int argc, char ** argv) {
  v2sw a = {1,2,3,4};
  v2sw b = {5,6,7,8};
  v2sw c = a;

  int i;

  for (i=0;i<argc;i++)
    c += b;

  printf("%d %d %d %d\n",c[0],c[1],c[2],c[3]);

  return 0;
}

When I compile this to bytecode with clang and then invoke llc with -march=mips, everything goes fine and the code is scalarized. But when I try it on my backend, I get
LLVM ERROR: Cannot select: 0xa5e2730: v4i16 = BUILD_VECTOR 0xa5e29d8, 0xa5e2378, 0xa5e21e0, 0xa5e2158 [ID=14]
  0xa5e29d8: i32 = Constant<1> [ID=11]
  0xa5e2378: i32 = Constant<2> [ID=10]
  0xa5e21e0: i32 = Constant<3> [ID=9]
  0xa5e2158: i32 = Constant<4> [ID=8]

Now I've looked at the Mips backend and other backends in detail, but I didn't find anything that would force the scalarization in there (like references to v64, v4i16, BUILD_VECTOR, ...), so I am clueless ...

Thanks,

Bjorn De Sutter
Computer Systems Lab
Ghent University