[llvm] [AMDGPU][True16] Generate correct reg size for reg_sequence16 in wmma src mod select (PR #187629)

[Guo Chen via llvm-commits]

================
@@ -3752,22 +3753,42 @@ static MachineSDNode *buildRegSequence16(SmallVectorImpl<SDValue> &Elts,
     if (isExtractHiElt(Elts[i + 1], HiSrc) && LoSrc == HiSrc) {
       PackedElts.push_back(HiSrc);
     } else {
+      if (Subtarget->useRealTrue16Insts()) {
+        // FIXME-TRUE16. use reg_sequence to replace v_perm_b32
----------------
broxigarchen wrote:

Updated the comment to make it more clear.

Factor it to be a seperate function. Does this looks good to you?

https://github.com/llvm/llvm-project/pull/187629