Missing and Wrong Data Layout Transformation Checks

[andrej]

The compiler currently (1) rejects some data layout transformations that are legal and (2) fails to reject some data layout transformations that are not realizable in hardware (i..e, sub-32-bit accesses).

For (1), this check, introduced in #2435, effectively disables use of the lowest-dimension stride:

mlir-aie/lib/Dialect/AIE/IR/AIEDialect.cpp

Line 1809 in d98dd25

if (getBufferElementTypeWidthInBytes() > 4 && dims->back().getStride() != 1)

I don't know the background of why it was introduced, but I believe it effectively makes the lowest-dimension stride unusable.

For (2), as long as the lowest-dimension stride is 1 for smaller data types (e.g. int8) the compiler currently does not reject smaller-than-4-bytes transfers. Something like [<size=3, stride=4>, <size=2, stride=1>] should be rejected if the data type is int8, since the smallest unit of transfer here is 16 bits.

We have this function used elsewhere:

mlir-aie/lib/Dialect/AIEX/IR/AIEXDialect.cpp

Line 171 in d98dd25

AIEX::verifyStridesWraps(mlir::Operation *forOp,

I think we should try to factor that out and use that for verification elsewhere as well. It probably contains mistakes too but not the above two.