[llvm] [AArch64] Don't use LowerToPredicatedOp to shufflevector -> SVE lowerings (PR #140713)

[Benjamin Maxwell via llvm-commits]

MacDue wrote:

> Can you point me to where that happens? I'm looking at the implementation of `LowerToPredicatedOp` and I don't see it trying to look through the bitcasts?

The value "`Op`" passed to `LowerToPredicatedOp` is not used as an operand of the node `LowerToPredicatedOp` returns, instead `LowerToPredicatedOp` turns the operand into the new node. So when you create a new BITCAST node and then immediately pass that node to `LowerToPredicatedOp` the result is not a new node that uses that BITCAST, it's a node that takes all its operands from the BITCAST + some predicate and replaces the opcode, the BITCAST node is unused. 

https://github.com/llvm/llvm-project/pull/140713