[llvm] [LV] Use SCEV to compute final value of complex induction variables (PR #195059)

[David Sherwood via llvm-commits]

================
@@ -9,77 +9,61 @@ define i64 @reduction(i64 %arg) #0 {
 ; CHECK:       [[VECTOR_PH]]:
 ; CHECK-NEXT:    br i1 false, label %[[VEC_EPILOG_PH:.*]], label %[[VECTOR_PH1:.*]]
 ; CHECK:       [[VECTOR_PH1]]:
-; CHECK-NEXT:    [[BROADCAST_SPLATINSERT:%.*]] = insertelement <4 x i64> poison, i64 [[ARG]], i64 0
-; CHECK-NEXT:    [[BROADCAST_SPLAT:%.*]] = shufflevector <4 x i64> [[BROADCAST_SPLATINSERT]], <4 x i64> poison, <4 x i32> zeroinitializer
 ; CHECK-NEXT:    br label %[[VECTOR_BODY:.*]]
 ; CHECK:       [[VECTOR_BODY]]:
 ; CHECK-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, %[[VECTOR_PH1]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]
-; CHECK-NEXT:    [[VEC_IND:%.*]] = phi <4 x i64> [ <i64 0, i64 1, i64 2, i64 3>, %[[VECTOR_PH1]] ], [ [[VEC_IND_NEXT:%.*]], %[[VECTOR_BODY]] ]
 ; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi <4 x i32> [ zeroinitializer, %[[VECTOR_PH1]] ], [ [[TMP5:%.*]], %[[VECTOR_BODY]] ]
 ; CHECK-NEXT:    [[VEC_PHI2:%.*]] = phi <4 x i32> [ zeroinitializer, %[[VECTOR_PH1]] ], [ [[TMP1:%.*]], %[[VECTOR_BODY]] ]
 ; CHECK-NEXT:    [[VEC_PHI4:%.*]] = phi <4 x i32> [ zeroinitializer, %[[VECTOR_PH1]] ], [ [[TMP16:%.*]], %[[VECTOR_BODY]] ]
 ; CHECK-NEXT:    [[VEC_PHI3:%.*]] = phi <4 x i32> [ zeroinitializer, %[[VECTOR_PH1]] ], [ [[TMP17:%.*]], %[[VECTOR_BODY]] ]
-; CHECK-NEXT:    [[STEP_ADD:%.*]] = add nuw <4 x i64> [[VEC_IND]], splat (i64 4)
-; CHECK-NEXT:    [[STEP_ADD_2:%.*]] = add nuw <4 x i64> [[STEP_ADD]], splat (i64 4)
-; CHECK-NEXT:    [[STEP_ADD_3:%.*]] = add nuw <4 x i64> [[STEP_ADD_2]], splat (i64 4)
 ; CHECK-NEXT:    [[TMP5]] = or <4 x i32> [[VEC_PHI]], splat (i32 1)
 ; CHECK-NEXT:    [[TMP1]] = or <4 x i32> [[VEC_PHI2]], splat (i32 1)
 ; CHECK-NEXT:    [[TMP16]] = or <4 x i32> [[VEC_PHI4]], splat (i32 1)
 ; CHECK-NEXT:    [[TMP17]] = or <4 x i32> [[VEC_PHI3]], splat (i32 1)
 ; CHECK-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 16
-; CHECK-NEXT:    [[VEC_IND_NEXT]] = add nsw <4 x i64> [[STEP_ADD_3]], splat (i64 4)
 ; CHECK-NEXT:    [[TMP6:%.*]] = icmp eq i64 [[INDEX_NEXT]], 96
 ; CHECK-NEXT:    br i1 [[TMP6]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]
 ; CHECK:       [[MIDDLE_BLOCK]]:
-; CHECK-NEXT:    [[TMP18:%.*]] = add nsw <4 x i64> [[STEP_ADD_3]], splat (i64 1)
-; CHECK-NEXT:    [[TMP19:%.*]] = mul nsw <4 x i64> [[TMP18]], [[BROADCAST_SPLAT]]
 ; CHECK-NEXT:    [[BIN_RDX1:%.*]] = or <4 x i32> [[TMP1]], [[TMP5]]
 ; CHECK-NEXT:    [[BIN_RDX4:%.*]] = or <4 x i32> [[TMP16]], [[BIN_RDX1]]
 ; CHECK-NEXT:    [[BIN_RDX5:%.*]] = or <4 x i32> [[TMP17]], [[BIN_RDX4]]
 ; CHECK-NEXT:    [[TMP7:%.*]] = call i32 @llvm.vector.reduce.or.v4i32(<4 x i32> [[BIN_RDX5]])
-; CHECK-NEXT:    [[TMP8:%.*]] = extractelement <4 x i64> [[TMP19]], i64 3
+; CHECK-NEXT:    [[TMP8:%.*]] = mul i64 95, [[ARG]]
----------------
david-arm wrote:

Also, in the original vector code prior to this PR the value was calculated as:

```
  [[TMP18:%.*]] = add nsw <4 x i64> [[STEP_ADD_3]], splat (i64 1)
  [[TMP19:%.*]] = mul nsw <4 x i64> [[TMP18]], [[BROADCAST_SPLAT]]
```

where it looks like we are properly accounting for the increment by adding a splat of 1.

https://github.com/llvm/llvm-project/pull/195059