[llvm] [VPlan] Run removeDeadRecipes early. (PR #190191)

[Florian Hahn via llvm-commits]

================
@@ -1058,9 +1058,14 @@ define void @pointer_iv_non_uniform_0(ptr %a, i64 %n) {
 ; INTER-NEXT:    [[STRIDED_VEC:%.*]] = shufflevector <16 x i32> [[WIDE_VEC]], <16 x i32> poison, <4 x i32> <i32 0, i32 4, i32 8, i32 12>
 ; INTER-NEXT:    [[STRIDED_VEC4:%.*]] = shufflevector <16 x i32> [[WIDE_VEC]], <16 x i32> poison, <4 x i32> <i32 1, i32 5, i32 9, i32 13>
 ; INTER-NEXT:    [[TMP7:%.*]] = getelementptr inbounds i32, ptr [[NEXT_GEP]], i32 4
-; INTER-NEXT:    [[WIDE_VEC5:%.*]] = load <16 x i32>, ptr [[TMP7]], align 8
-; INTER-NEXT:    [[STRIDED_VEC6:%.*]] = shufflevector <16 x i32> [[WIDE_VEC5]], <16 x i32> poison, <4 x i32> <i32 0, i32 4, i32 8, i32 12>
-; INTER-NEXT:    [[STRIDED_VEC7:%.*]] = shufflevector <16 x i32> [[WIDE_VEC5]], <16 x i32> poison, <4 x i32> <i32 1, i32 5, i32 9, i32 13>
+; INTER-NEXT:    [[TMP28:%.*]] = getelementptr inbounds i32, ptr [[NEXT_GEP1]], i32 4
+; INTER-NEXT:    [[TMP29:%.*]] = getelementptr inbounds i32, ptr [[NEXT_GEP2]], i32 4
+; INTER-NEXT:    [[TMP30:%.*]] = getelementptr inbounds i32, ptr [[NEXT_GEP3]], i32 4
+; INTER-NEXT:    [[TMP31:%.*]] = insertelement <4 x ptr> poison, ptr [[TMP7]], i32 0
+; INTER-NEXT:    [[TMP32:%.*]] = insertelement <4 x ptr> [[TMP31]], ptr [[TMP28]], i32 1
+; INTER-NEXT:    [[TMP33:%.*]] = insertelement <4 x ptr> [[TMP32]], ptr [[TMP29]], i32 2
+; INTER-NEXT:    [[TMP34:%.*]] = insertelement <4 x ptr> [[TMP33]], ptr [[TMP30]], i32 3
+; INTER-NEXT:    [[STRIDED_VEC6:%.*]] = call <4 x i32> @llvm.masked.gather.v4i32.v4p0(<4 x ptr> align 8 [[TMP34]], <4 x i1> splat (i1 true), <4 x i32> poison)
----------------
fhahn wrote:

> Worth fixing to avoid landing a regression?

We could try, but I am not sure it is worth it? In the regular pipeline those loads would be removed earlier, and we would not form an interleave group in the first place, so this would be just fixing for artificial test cases I think

https://github.com/llvm/llvm-project/pull/190191