Execution engine crash on complex code using "sparse" encoding

[aartbik]

The execution engine crashes on the LLVM IR generated by MLIR lowering on IR that uses a "sparse" encoding of a complex32 buffer. I could narrow it down to the following reproducer:

module attributes {llvm.data_layout = ""} {

memref.global "private" constant @"__constant_32xcomplex_0" : memref<32xcomplex> =
sparse<[[1], [28], [31]],
[(1.000000e+00,0.000000e+00), (2.000000e+00,0.000000e+00), (3.000000e+00,0.000000e+00)]
> {alignment = 128 : i64}

llvm.func @entry() {
%0 = memref.get_global @"__constant_32xcomplex_0" : memref<32xcomplex>
llvm.return
}
}

[llvmbot]

@llvm/issue-subscribers-mlir-sparse

[llvmbot]

@llvm/issue-subscribers-mlir

[aartbik]

To reproduce:

mlir-opt --convert-memref-to-llvm repro.mlir | mlir-cpu-runner -e entry -entry-point-result=void

[aartbik]

This is blocking migration of sparse compiler to @matthias-springer 's excellent new bufferization strategy.

@bixia1 @PeimingLiu

[aartbik]

Looks like complex type is simply missing here ;-)

Attribute SparseElementsAttr::getZeroAttr() const {
  auto eltType = getElementType();

  // Handle floating point elements.
  if (eltType.isa<FloatType>())
    return FloatAttr::get(eltType, 0);

  // Handle string type.
  if (getValues().isa<DenseStringElementsAttr>())
    return StringAttr::get("", eltType);

  // Otherwise, this is an integer.
  return IntegerAttr::get(eltType, 0);
}

[llvmbot]

@llvm/issue-subscribers-mlir-core