[mlir][tensor] Refine the semantics of createPadHighOp (llvm#109667)

Refine `createPadHighOp` so that the output tensor is required to be
statically shaped. This is to prevent the current behaviour, which is
incorrect:

>  // If `type` has dynamic dimensions the padding width is set to zero.

The actual padding width should be set to: `%new_dim - %old_dim`, where
%new_dim` and `%old_dim` are defined via e.g. `tensor.dim` Op applied to
output and input tensors, respectively.

This PR is an attempt to clarify the semantics surrounding dynamic
shapes in preparation for adding support for scalable vectors to the
pack/unpack logic in Tensor/Linalg (dynamic shapes is what we use to
model scalable (*) sizes at the Tensor/MemRef level).

(*) Scalable as in Arm's Scalable Vector Extension (SVE)

mlir/include/mlir/Dialect/Tensor/Utils/Utils.h

@@ -14,10 +14,10 @@
 namespace mlir {
 namespace tensor {
 
-// Return a PadOp that pads `source` to `type` size where the static
-// sizes are assumed to be greater than the dynamic sizes. If `type` has dynamic
-// dimensions the padding width is set to zero. The op performs "high" padding
-// (i.e. it adds trailing padding values until the desired size is met).
+// Return a PadOp that pads `source` to `type` size. Output sizes (from `type`)
+// are assumed to be static and greater than the potentially dynamic input sizes
+// (from `source). The op performs "high" padding (i.e. it adds trailing padding
+// values until the desired size is met).
 PadOp createPadHighOp(RankedTensorType type, Value source, Value pad,
                       bool nofold, Location loc, OpBuilder &builder);
 

mlir/lib/Dialect/Tensor/Utils/Utils.cpp

@@ -24,12 +24,17 @@ using namespace mlir::tensor;
 PadOp mlir::tensor::createPadHighOp(RankedTensorType type, Value source,
                                     Value pad, bool nofold, Location loc,
                                     OpBuilder &b) {
+
+  // TODO: Either relax or turn this into a failure
+  assert(!ShapedType::isDynamicShape(type.getShape()) &&
+         "The output type is dynamic - that's not supported ATM.");
+
+  // Init "low" and "high" padding values ("low" is kept as is, "high" is
+  // computed below).
   SmallVector<OpFoldResult> low(type.getRank(), b.getIndexAttr(0));
   SmallVector<OpFoldResult> high(type.getRank(), b.getIndexAttr(0));
+
   for (const auto &en : enumerate(type.getShape())) {
-    // Pad only the static dimensions of the result tensor type.
-    if (ShapedType::isDynamic(en.value()))
-      continue;
     // Compute the padding width.
     AffineExpr d0;
     bindDims(b.getContext(), d0);