Dynamic vectorization

Strip-mine, then mask

Dynamic vectorization keeps the vector shape static and makes the trip count dynamic. The loop walks the array in vector-sized steps, and the final step is short:

%mask = vector.create_mask %bound : vector<[4]xi1>
%v = vector.transfer_read %A[%i], %pad, %mask
    {in_bounds = [false]} : memref<?xf32>, vector<[4]xf32>

The mask is how the source says which lanes are real. On RISC-V it can lower to a masked load, or to a length-bounded load where vl carries the count. Both are correct only if the mask and the length describe the same edge.

The length is an operand

vsetvli does not know the array bound. It computes vl = min(avl, VLMAX), where AVL is a value you put in a register:

sub     a0, s1, s0        # avl = n - i
vsetvli t0, a0, e32, m1, ta, ma
vle32.v v8, (a1)

The bound lives in a0, computed by ordinary scalar code. VLMAX comes from LMUL, VLEN and SEW; vl is the smaller of that and the remaining count. A loop that advances by VLMAX instead of the returned vl is wrong at the edge, and the error is a wrong result rather than a fault. This is the same subtraction a scalar loop performs, now feeding the vector configuration instead of a branch.

Check the mask op and the transfer syntax on your pinned release; vector.mask and the predicate form have moved more than once. vector.create_mask takes index operands.

Things that break

Using VLMAX as the step overruns the buffer when the available length was smaller. A mask computed from the tile index rather than the element count leaves the last lanes live. And a short final iteration under ta leaves the tail lanes unspecified, so an accumulator that survives the loop needs tu or an explicit merge. Confirm which policy your build emits before trusting a tail sum.