[llvm] [LV] Support strided memory accesses with a stride of -1 (PR #128718)

[Mel Chen via llvm-commits]

================
@@ -6132,6 +6174,16 @@ void LoopVectorizationCostModel::setCostBasedWideningDecision(ElementCount VF) {
                "Expected consecutive stride.");
         InstWidening Decision =
             ConsecutiveStride == 1 ? CM_Widen : CM_Widen_Reverse;
+        // Consider using strided load/store for consecutive reverse accesses to
+        // achieve more efficient memory operations.
+        if (ConsecutiveStride == -1 && stridedAccessCanBeWidened(&I, VF)) {
----------------
Mel-Chen wrote:

I'm still deciding between `if (Legal->isConsecutivePtr(ScalarTy, Ptr) != -1)` and `if (!Legal->isConsecutivePtr(ScalarTy, Ptr))`.
Since, in terms of feasibility, both stride 1 and -1 can be used for strided memory access—it’s just a matter of potentially higher cost.

https://github.com/llvm/llvm-project/pull/128718