[llvm] [LV] Use frozen start value for FindLastIV if needed. (PR #132691)

[Florian Hahn via llvm-commits]

================
@@ -6,6 +6,7 @@ define i64 @select_icmp_nuw_nsw(ptr %a, ptr %b, i64 %ii, i64 %n) {
 ; CHECK-SAME: ptr [[A:%.*]], ptr [[B:%.*]], i64 [[II:%.*]], i64 [[N:%.*]]) {
 ; CHECK-NEXT:  [[ENTRY:.*]]:
 ; CHECK-NEXT:    [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[N]], 4
+; CHECK-NEXT:    [[FR:%.*]] = freeze i64 [[II]]
----------------
fhahn wrote:

I checked a bit more and I think the only issue is with epilogue vectorization; Here's a proof where we go via the main vector loop to the scalar remainder loop https://alive2.llvm.org/ce/z/yFSBrq  and here where we go from the vector loop to the exit block https://alive2.llvm.org/ce/z/D6Ppzr.

When vectorizing the epilogue, freeze is needed for the start value used in the selects and the resume phi from the main vector loop:

* Both main and epilogue vector loops execute, go to exit block: https://alive2.llvm.org/ce/z/_TSvRr
* Both main and epilogue vector loops execute, go to scalar loop: https://alive2.llvm.org/ce/z/CsPj5v
* Only epilogue vector loop executes, go to exit block: https://alive2.llvm.org/ce/z/5XqkNV
* Only epilogue vector loop executes, go to scalar loop: https://alive2.llvm.org/ce/z/JUpqRN

The latter 2 show  requiring freezing the resume phi. That means we cannot freeze in the preheader. We could move the freeze to the main iteration count check, but that would be a bit fragile to find and other transforms can sink the freeze if needed.




https://github.com/llvm/llvm-project/pull/132691