[llvm] [VectorCombine] Add support for zext/sext/trunc to shuffleToIdentity (PR #92696)

[Ramkumar Ramachandra via llvm-commits]

================
@@ -754,10 +720,8 @@ define void @trunc(<8 x i64> %a, <8 x i64> %b, ptr %p) {
 
 define <4 x i64> @zext_chain(<4 x i16> %x) {
----------------
artagnon wrote:

Can you kindly verify this? I'm mainly worried about some underlying bug in the algorithm, which could result in correctness issues as opposed to missed optimizations. It seems very strange to me that a single add, but not a chain of adds, is handled. If multiple uses aren't handled, even a single add with identical operands shouldn't be reduced. If a chain of adds and chain of zexts are handled, a combination of the two should be handled automatically. At least that's how my (now abandoned) algorithm worked.

https://github.com/llvm/llvm-project/pull/92696