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

[Florian Hahn via llvm-commits]

================
@@ -74,15 +74,15 @@ define void @vector_reverse_i64(ptr nocapture noundef writeonly %A, ptr nocaptur
 ; CHECK-NEXT:      EMIT vp<[[CAN_IV:%.+]]> = CANONICAL-INDUCTION
 ; CHECK-NEXT:      vp<[[DEV_IV:%.+]]> = DERIVED-IV ir<%n> + vp<[[CAN_IV]]> * ir<-1>
 ; CHECK-NEXT:      vp<[[STEPS:%.+]]> = SCALAR-STEPS vp<[[DEV_IV]]>, ir<-1>
-; CHECK-NEXT:      CLONE ir<%i.0> = add nsw vp<[[STEPS]]>, ir<-1>
-; CHECK-NEXT:      CLONE ir<%idxprom> = zext ir<%i.0>
-; CHECK-NEXT:      CLONE ir<%arrayidx> = getelementptr inbounds ir<%B>, ir<%idxprom>
-; CHECK-NEXT:      vp<[[VEC_PTR:%.+]]> = vector-end-pointer inbounds ir<%arrayidx>, vp<[[VF]]>
-; CHECK-NEXT:      WIDEN ir<%1> = load vp<[[VEC_PTR]]>
-; CHECK-NEXT:      WIDEN ir<%add9> = add ir<%1>, ir<1>
-; CHECK-NEXT:      CLONE ir<%arrayidx3> = getelementptr inbounds ir<%A>, ir<%idxprom>
-; CHECK-NEXT:      vp<[[VEC_PTR2:%.+]]> = vector-end-pointer inbounds ir<%arrayidx3>, vp<[[VF]]>
-; CHECK-NEXT:      WIDEN store vp<[[VEC_PTR2]]>, ir<%add9>
+; CHECK-NEXT:      CLONE ir<[[IDX:%.+]]> = add nsw vp<[[STEPS]]>, ir<-1>
+; CHECK-NEXT:      CLONE ir<[[ZEXT_IDX:%.+]]> = zext ir<[[IDX]]>
+; CHECK-NEXT:      CLONE ir<[[LD_IDX:%.+]]> = getelementptr inbounds ir<%B>, ir<[[ZEXT_IDX]]>
+; CHECK-NEXT:      vp<[[LD_PTR:%.+]]> = vector-pointer ir<[[LD_IDX]]>
+; CHECK-NEXT:      WIDEN ir<[[LD:%.+]]> = load vp<[[LD_PTR]]>, stride = ir<-4>, runtimeVF = vp<[[VF]]>
----------------
fhahn wrote:

I am a bit surprised why this strides load with stride of 1 is more profitable than a consecutive load?  

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