[llvm] [VPlan] Extend CSE to eliminate GEPs (PR #156699)

[Ramkumar Ramachandra via llvm-commits]

================
@@ -136,7 +136,7 @@ define void @vectorize_without_optsize(ptr %p, i32 %x, i64 %n) {
 ; DEFAULT:       [[VECTOR_BODY]]:
 ; DEFAULT-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]
 ; DEFAULT-NEXT:    [[TMP1:%.*]] = getelementptr inbounds i32, ptr [[P]], i64 [[INDEX]]
-; DEFAULT-NEXT:    [[TMP3:%.*]] = getelementptr inbounds i32, ptr [[TMP1]], i32 4
+; DEFAULT-NEXT:    [[TMP3:%.*]] = getelementptr i32, ptr [[TMP1]], i32 4
----------------
artagnon wrote:

Unfortunately, I don't have a cycle-accurate way of verifying that there are no regressions: timing with perf natively on X86 is far too noisy to be useful :(

https://github.com/llvm/llvm-project/pull/156699