In the buildShuffleVector method, there is a restriction that the
input vectors must be larger than the mask. However, this is not
the definition of a ShuffleVector instruction that is used inside
of our test suite.

For example:
shuffle_concat_1 in combine_shuffle_vector.mir: 4xs8 -> 16xs8
v3s8_crash in legalize_insert_vector_elt: 3xs8 -> 12xs8
shuffle_vector_to_concat_vector_45670123 in prelegalizercombiner-shuffle-vector: 4xs32 -> 12xs32

…iner

… sequences into G_SHUFFLE_VECTOR

This combine tries to find all the build vectors whose source elements all originate from a G_EXTRACT_VECTOR_ELT from one or two donor vectors. One example where this may happen is for AI chips where there are a lot of matrix multiplications. Typically there vectors are disected and then  rearranged into the right transformation.
E.g.
  %donor1(<2 x s32>) = COPY $d0
  %donor2(<2 x s32>) = COPY $d1
  %ext1 = G_EXTRACT_VECTOR_ELT %donor1, 0
  %ext2 = G_EXTRACT_VECTOR_ELT %donor1, 1
  %ext3 = G_EXTRACT_VECTOR_ELT %donor2, 0
  %ext4 = G_EXTRACT_VECTOR_ELT %donor2, 1
  %vector = G_BUILD_VECTOR %ext1, %ext2, %ext3, %ext4
  ==>
  replace with:  %vector = G_SHUFFLE_VECTOR %donor1, %donor2, shufflemask(0, 1, 2, 3)