[llvm] [VPlan] Fix packed replication of struct types (PR #160274)

[Luke Lau via llvm-commits]

================
@@ -453,6 +453,221 @@ exit:
   ret void
 }
 
+define void @struct_return_2xf32_replicate_predicated(ptr %a) {
+; VF4-LABEL: define void @struct_return_2xf32_replicate_predicated(
+; VF4-SAME: ptr [[A:%.*]]) {
+; VF4-NEXT:  [[ENTRY:.*:]]
+; VF4-NEXT:    br label %[[VECTOR_PH:.*]]
+; VF4:       [[VECTOR_PH]]:
+; VF4-NEXT:    br label %[[VECTOR_BODY:.*]]
+; VF4:       [[VECTOR_BODY]]:
+; VF4-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[PRED_STORE_CONTINUE12:.*]] ]
+; VF4-NEXT:    [[TMP0:%.*]] = getelementptr inbounds float, ptr [[A]], i64 [[INDEX]]
+; VF4-NEXT:    [[WIDE_LOAD:%.*]] = load <4 x float>, ptr [[TMP0]], align 8
+; VF4-NEXT:    [[TMP1:%.*]] = fcmp ogt <4 x float> [[WIDE_LOAD]], zeroinitializer
+; VF4-NEXT:    [[TMP2:%.*]] = extractelement <4 x i1> [[TMP1]], i32 0
+; VF4-NEXT:    br i1 [[TMP2]], label %[[PRED_CALL_IF:.*]], label %[[PRED_CALL_CONTINUE:.*]]
+; VF4:       [[PRED_CALL_IF]]:
+; VF4-NEXT:    [[TMP3:%.*]] = extractelement <4 x float> [[WIDE_LOAD]], i32 0
+; VF4-NEXT:    [[TMP4:%.*]] = tail call { float, float } @fn2(float [[TMP3]]) #[[ATTR3:[0-9]+]]
+; VF4-NEXT:    [[TMP5:%.*]] = extractvalue { float, float } [[TMP4]], 0
+; VF4-NEXT:    [[TMP6:%.*]] = insertelement <4 x float> poison, float [[TMP5]], i32 0
+; VF4-NEXT:    [[TMP7:%.*]] = insertvalue { <4 x float>, <4 x float> } poison, <4 x float> [[TMP6]], 0
+; VF4-NEXT:    [[TMP8:%.*]] = extractvalue { float, float } [[TMP4]], 1
+; VF4-NEXT:    [[TMP9:%.*]] = extractvalue { <4 x float>, <4 x float> } [[TMP7]], 1
+; VF4-NEXT:    [[TMP10:%.*]] = insertelement <4 x float> [[TMP9]], float [[TMP8]], i32 0
+; VF4-NEXT:    [[TMP11:%.*]] = insertvalue { <4 x float>, <4 x float> } [[TMP7]], <4 x float> [[TMP10]], 1
+; VF4-NEXT:    br label %[[PRED_CALL_CONTINUE]]
+; VF4:       [[PRED_CALL_CONTINUE]]:
+; VF4-NEXT:    [[TMP12:%.*]] = phi { <4 x float>, <4 x float> } [ poison, %[[VECTOR_BODY]] ], [ [[TMP7]], %[[PRED_CALL_IF]] ]
+; VF4-NEXT:    [[TMP13:%.*]] = extractelement <4 x i1> [[TMP1]], i32 1
+; VF4-NEXT:    br i1 [[TMP13]], label %[[PRED_CALL_IF1:.*]], label %[[PRED_CALL_CONTINUE2:.*]]
+; VF4:       [[PRED_CALL_IF1]]:
+; VF4-NEXT:    [[TMP14:%.*]] = extractelement <4 x float> [[WIDE_LOAD]], i32 1
+; VF4-NEXT:    [[TMP15:%.*]] = tail call { float, float } @fn2(float [[TMP14]]) #[[ATTR3]]
+; VF4-NEXT:    [[TMP16:%.*]] = extractvalue { float, float } [[TMP15]], 0
+; VF4-NEXT:    [[TMP17:%.*]] = extractvalue { <4 x float>, <4 x float> } [[TMP12]], 0
+; VF4-NEXT:    [[TMP18:%.*]] = insertelement <4 x float> [[TMP17]], float [[TMP16]], i32 1
+; VF4-NEXT:    [[TMP19:%.*]] = insertvalue { <4 x float>, <4 x float> } [[TMP12]], <4 x float> [[TMP18]], 0
+; VF4-NEXT:    [[TMP20:%.*]] = extractvalue { float, float } [[TMP15]], 1
+; VF4-NEXT:    [[TMP21:%.*]] = extractvalue { <4 x float>, <4 x float> } [[TMP19]], 1
+; VF4-NEXT:    [[TMP22:%.*]] = insertelement <4 x float> [[TMP21]], float [[TMP20]], i32 1
+; VF4-NEXT:    [[TMP23:%.*]] = insertvalue { <4 x float>, <4 x float> } [[TMP19]], <4 x float> [[TMP22]], 1
----------------
lukel97 wrote:

I don't think it is. All the crashes happened in test files with -force-vector-width. It just so happened that changing the predication discount with BFI changed the scalarization decision which led to this in struct-return.ll

https://github.com/llvm/llvm-project/pull/160274