Error when slicing MemRef with memory_space

[Ritsuka314]

MemRefT = MemRef.get((DYNAMIC,), F32, memory_space=GPU_AddrSpace.Global)
@compile(auto_build=False)
def f(m: MemRefT):
    n = m[:]

gives error

mlir._mlir_libs._site_initialize.<locals>.MLIRError: Verification failed:
error: unknown: different memory spaces specified for
base memref type 'memref<?xf32, #gpu.address_space<global>>' and
subview memref type 'memref<?xf32, strided<[?], offset: ?>>'
 note: unknown: see current operation: %6 = "memref.subview"(%arg0, %2, %5, %3) <{operandSegmentSizes = array<i32: 1, 1, 1, 1>, static_offsets = array<i64: -9223372036854775808>, static_sizes = array<i64: -9223372036854775808>, static_strides = array<i64: -9223372036854775808>}> : (memref<?xf32, #gpu.address_space<global>>, index, index, index) -> memref<?xf32, strided<[?], offset: ?>>

[Balint-R]

Ah, right. I think in general, it might be useful to add a new "constructor" to memref types (the current "constructor" being class_factory) that takes in an existing memref type and a list of fields to change, and returns a new memref type with those fields changed.

E.g.

@classmethod
def get_similar_type(cls, *, element_type=None, shape=None, ...):

#96 is relevant. It already implements some of this logic.

One issue with this is that None is a valid value for some of the fields (e.g. strides, memory_space), so can't always use that as the default value in the function definition :/.