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

[Alexey Bataev via llvm-commits]

================
@@ -2702,12 +2709,60 @@ struct VPWidenLoadEVLRecipe final : public VPWidenMemoryRecipe, public VPValue {
   bool onlyFirstLaneUsed(const VPValue *Op) const override {
     assert(is_contained(operands(), Op) &&
            "Op must be an operand of the recipe");
-    // Widened loads only demand the first lane of EVL and consecutive loads
-    // only demand the first lane of their address.
+    // Widened loads only demand the first lane of EVL and consecutive/strided
+    // loads only demand the first lane of their address.
     return Op == getEVL() || (Op == getAddr() && isConsecutive());
   }
 };
 
+/// A recipe for strided load operations, using the base address, stride, and an
+/// optional mask.
----------------
alexey-bataev wrote:

Possible in its current state, but not sure it is possible in general. Some (potential) strided memory recipes are currently represented as masked gathers, which require widened pointers. Strided/interleaved mem accesses do not require widened (vector) pointers, so it may affect the cost significantly. I think we need to teach the whole loop vectorizer the strided mem accesses pattern (no matter if the stride is constant, as it is right now, or runtime defined)

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