[VPlan] Fix handling of ReductionStartVector for rdxs when unrolling.

Update handling of ReductionStartVector in VPlanUnroll for partial
reductions. The new code makes sure all parts are properly set to the
cloned ReductionStartVector.

Fixes a mis-compile reported for
#142290.

Author: fhahn

llvm/lib/Transforms/Vectorize/VPlanUnroll.cpp

@@ -231,14 +231,19 @@ void UnrollState::unrollHeaderPHIByUF(VPHeaderPHIRecipe *R,
       if (auto *VPI = dyn_cast<VPInstruction>(RdxPhi->getStartValue())) {
         assert(VPI->getOpcode() == VPInstruction::ReductionStartVector &&
                "unexpected start VPInstruction");
+        if (Part != 1)
+          continue;
+        VPValue *StartV;
         if (match(VPI->getOperand(2), m_SpecificInt(1))) {
-          Copy->setOperand(0, VPI->getOperand(1));
-        } else if (Part == 1) {
+          StartV = VPI->getOperand(1);
+        } else {
           auto *C = VPI->clone();
           C->setOperand(0, C->getOperand(1));
           C->insertAfter(VPI);
-          addUniformForAllParts(C);
+          StartV = C;
         }
+        for (unsigned Part = 1; Part != UF; ++Part)
+          VPV2Parts[VPI][Part - 1] = StartV;
       }
       Copy->addOperand(getConstantVPV(Part));
     } else {

llvm/test/Transforms/LoopVectorize/AArch64/partial-reduce-interleave.ll

@@ -18,7 +18,7 @@ define i32 @partial_reduce_with_non_constant_start_value(ptr %src, i32 %rdx.star
 ; IC2:       [[VECTOR_BODY]]:
 ; IC2-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]
 ; IC2-NEXT:    [[VEC_PHI:%.*]] = phi <4 x i32> [ [[TMP0]], %[[VECTOR_PH]] ], [ [[PARTIAL_REDUCE:%.*]], %[[VECTOR_BODY]] ]
-; IC2-NEXT:    [[VEC_PHI1:%.*]] = phi <4 x i32> [ [[TMP0]], %[[VECTOR_PH]] ], [ [[PARTIAL_REDUCE3:%.*]], %[[VECTOR_BODY]] ]
+; IC2-NEXT:    [[VEC_PHI1:%.*]] = phi <4 x i32> [ zeroinitializer, %[[VECTOR_PH]] ], [ [[PARTIAL_REDUCE3:%.*]], %[[VECTOR_BODY]] ]
 ; IC2-NEXT:    [[TMP1:%.*]] = getelementptr inbounds i8, ptr [[SRC]], i64 [[INDEX]]
 ; IC2-NEXT:    [[TMP2:%.*]] = getelementptr inbounds i8, ptr [[TMP1]], i32 0
 ; IC2-NEXT:    [[TMP3:%.*]] = getelementptr inbounds i8, ptr [[TMP1]], i32 16
@@ -70,9 +70,9 @@ define i32 @partial_reduce_with_non_constant_start_value(ptr %src, i32 %rdx.star
 ; IC4:       [[VECTOR_BODY]]:
 ; IC4-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]
 ; IC4-NEXT:    [[VEC_PHI:%.*]] = phi <4 x i32> [ [[TMP0]], %[[VECTOR_PH]] ], [ [[PARTIAL_REDUCE:%.*]], %[[VECTOR_BODY]] ]
-; IC4-NEXT:    [[VEC_PHI1:%.*]] = phi <4 x i32> [ [[TMP0]], %[[VECTOR_PH]] ], [ [[PARTIAL_REDUCE7:%.*]], %[[VECTOR_BODY]] ]
-; IC4-NEXT:    [[VEC_PHI2:%.*]] = phi <4 x i32> [ [[TMP0]], %[[VECTOR_PH]] ], [ [[PARTIAL_REDUCE8:%.*]], %[[VECTOR_BODY]] ]
-; IC4-NEXT:    [[VEC_PHI3:%.*]] = phi <4 x i32> [ [[TMP0]], %[[VECTOR_PH]] ], [ [[PARTIAL_REDUCE9:%.*]], %[[VECTOR_BODY]] ]
+; IC4-NEXT:    [[VEC_PHI1:%.*]] = phi <4 x i32> [ zeroinitializer, %[[VECTOR_PH]] ], [ [[PARTIAL_REDUCE7:%.*]], %[[VECTOR_BODY]] ]
+; IC4-NEXT:    [[VEC_PHI2:%.*]] = phi <4 x i32> [ zeroinitializer, %[[VECTOR_PH]] ], [ [[PARTIAL_REDUCE8:%.*]], %[[VECTOR_BODY]] ]
+; IC4-NEXT:    [[VEC_PHI3:%.*]] = phi <4 x i32> [ zeroinitializer, %[[VECTOR_PH]] ], [ [[PARTIAL_REDUCE9:%.*]], %[[VECTOR_BODY]] ]
 ; IC4-NEXT:    [[TMP1:%.*]] = getelementptr inbounds i8, ptr [[SRC]], i64 [[INDEX]]
 ; IC4-NEXT:    [[TMP2:%.*]] = getelementptr inbounds i8, ptr [[TMP1]], i32 0
 ; IC4-NEXT:    [[TMP3:%.*]] = getelementptr inbounds i8, ptr [[TMP1]], i32 16