[llvm] [InstCombine] foldVecExtTruncToExtElt - extend to handle trunc(lshr(extractelement(x,c1),c2)) -> extractelement(bitcast(x),c3) patterns. (PR #109689)

[Simon Pilgrim via llvm-commits]

================
@@ -469,10 +473,23 @@ static Instruction *foldVecExtTruncToExtElt(TruncInst &Trunc,
   uint64_t BitCastNumElts = VecElts.getKnownMinValue() * TruncRatio;
   uint64_t VecOpIdx = Cst->getZExtValue();
   uint64_t NewIdx = IC.getDataLayout().isBigEndian()
-                        ? (VecOpIdx + 1) * TruncRatio - 1
+                        ? (VecOpIdx * TruncRatio) + (TruncRatio - 1)
----------------
RKSimon wrote:

Yes, in an earlier draft I was trying to get the DAG/InstCombine implementations to match - I've undone the change

https://github.com/llvm/llvm-project/pull/109689