[VPlan] Limit ExtractLastElem fold to recipes guaranteed single-scalar.

vputils::isSingleScalar(A) may return true to recipes that produce only
a single scalar value, but they could still end up as vector
instruction, because the recipe could not be converted to a
single-scalar VPInstruction/VPReplicateRecipe.

For now, only apply the fold for recipes guaranteed to produce a single
value, i.e. single-scalar VPInstructions and VPReplicateRecipes.

Fixes #158319.

Author: fhahn

llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp

@@ -1272,9 +1272,11 @@ static void simplifyRecipe(VPRecipeBase &R, VPTypeAnalysis &TypeInfo) {
 
   if (match(Def,
             m_VPInstruction<VPInstruction::ExtractLastElement>(m_VPValue(A))) &&
-      vputils::isSingleScalar(A) && all_of(A->users(), [Def, A](VPUser *U) {
-        return U->usesScalars(A) || Def == U;
-      })) {
+      ((isa<VPInstruction>(A) && vputils::isSingleScalar(A)) ||
+       (isa<VPReplicateRecipe>(A) &&
+        cast<VPReplicateRecipe>(A)->isSingleScalar())) &&
+      all_of(A->users(),
+             [Def, A](VPUser *U) { return U->usesScalars(A) || Def == U; })) {
     return Def->replaceAllUsesWith(A);
   }
 }

llvm/test/Transforms/LoopVectorize/first-order-recurrence-with-uniform-ops.ll

@@ -0,0 +1,248 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --check-globals none --version 6
+; RUN: opt -passes=loop-vectorize -force-vector-width=4 -force-vector-interleave=2 -S %s | FileCheck %s --check-prefix=UNROLL-NO-IC
+; RUN: opt -passes=loop-vectorize -force-vector-width=1 -force-vector-interleave=2 -S %s | FileCheck %s --check-prefix=UNROLL-NO-VF
+; RUN: opt -passes=loop-vectorize -force-vector-width=4 -force-vector-interleave=1 -S %s | FileCheck %s --check-prefix=SINK-AFTER
+
+; Test case for https://github.com/llvm/llvm-project/issues/95520.
+define i32 @recurence_uniform_load(ptr %src, ptr noalias %dst, i64 %n) {
+; UNROLL-NO-IC-LABEL: define i32 @recurence_uniform_load(
+; UNROLL-NO-IC-SAME: ptr [[SRC:%.*]], ptr noalias [[DST:%.*]], i64 [[N:%.*]]) {
+; UNROLL-NO-IC-NEXT:  [[ENTRY:.*]]:
+; UNROLL-NO-IC-NEXT:    [[TMP0:%.*]] = add i64 [[N]], 1
+; UNROLL-NO-IC-NEXT:    [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[TMP0]], 8
+; UNROLL-NO-IC-NEXT:    br i1 [[MIN_ITERS_CHECK]], label %[[SCALAR_PH:.*]], label %[[VECTOR_PH:.*]]
+; UNROLL-NO-IC:       [[VECTOR_PH]]:
+; UNROLL-NO-IC-NEXT:    [[N_MOD_VF:%.*]] = urem i64 [[TMP0]], 8
+; UNROLL-NO-IC-NEXT:    [[N_VEC:%.*]] = sub i64 [[TMP0]], [[N_MOD_VF]]
+; UNROLL-NO-IC-NEXT:    br label %[[VECTOR_BODY:.*]]
+; UNROLL-NO-IC:       [[VECTOR_BODY]]:
+; UNROLL-NO-IC-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]
+; UNROLL-NO-IC-NEXT:    [[TMP1:%.*]] = load i32, ptr [[SRC]], align 4
+; UNROLL-NO-IC-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 8
+; UNROLL-NO-IC-NEXT:    [[TMP2:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
+; UNROLL-NO-IC-NEXT:    br i1 [[TMP2]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]
+; UNROLL-NO-IC:       [[MIDDLE_BLOCK]]:
+; UNROLL-NO-IC-NEXT:    [[CMP_N:%.*]] = icmp eq i64 [[TMP0]], [[N_VEC]]
+; UNROLL-NO-IC-NEXT:    br i1 [[CMP_N]], label %[[EXIT:.*]], label %[[SCALAR_PH]]
+; UNROLL-NO-IC:       [[SCALAR_PH]]:
+; UNROLL-NO-IC-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC]], %[[MIDDLE_BLOCK]] ], [ 0, %[[ENTRY]] ]
+; UNROLL-NO-IC-NEXT:    [[SCALAR_RECUR_INIT:%.*]] = phi i32 [ [[TMP1]], %[[MIDDLE_BLOCK]] ], [ 0, %[[ENTRY]] ]
+; UNROLL-NO-IC-NEXT:    br label %[[LOOP:.*]]
+; UNROLL-NO-IC:       [[LOOP]]:
+; UNROLL-NO-IC-NEXT:    [[PHI:%.*]] = phi i64 [ [[BC_RESUME_VAL]], %[[SCALAR_PH]] ], [ [[ADD:%.*]], %[[LOOP]] ]
+; UNROLL-NO-IC-NEXT:    [[RECUR:%.*]] = phi i32 [ [[SCALAR_RECUR_INIT]], %[[SCALAR_PH]] ], [ [[LOAD:%.*]], %[[LOOP]] ]
+; UNROLL-NO-IC-NEXT:    [[ADD]] = add i64 [[PHI]], 1
+; UNROLL-NO-IC-NEXT:    [[LOAD]] = load i32, ptr [[SRC]], align 4
+; UNROLL-NO-IC-NEXT:    [[ICMP:%.*]] = icmp ult i64 [[PHI]], [[N]]
+; UNROLL-NO-IC-NEXT:    br i1 [[ICMP]], label %[[LOOP]], label %[[EXIT]], !llvm.loop [[LOOP3:![0-9]+]]
+; UNROLL-NO-IC:       [[EXIT]]:
+; UNROLL-NO-IC-NEXT:    [[RECUR_LCSSA:%.*]] = phi i32 [ [[RECUR]], %[[LOOP]] ], [ [[TMP1]], %[[MIDDLE_BLOCK]] ]
+; UNROLL-NO-IC-NEXT:    ret i32 [[RECUR_LCSSA]]
+;
+; UNROLL-NO-VF-LABEL: define i32 @recurence_uniform_load(
+; UNROLL-NO-VF-SAME: ptr [[SRC:%.*]], ptr noalias [[DST:%.*]], i64 [[N:%.*]]) {
+; UNROLL-NO-VF-NEXT:  [[ENTRY:.*]]:
+; UNROLL-NO-VF-NEXT:    [[TMP1:%.*]] = add i64 [[N]], 1
+; UNROLL-NO-VF-NEXT:    [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[TMP1]], 2
+; UNROLL-NO-VF-NEXT:    br i1 [[MIN_ITERS_CHECK]], label %[[SCALAR_PH:.*]], label %[[VECTOR_PH:.*]]
+; UNROLL-NO-VF:       [[VECTOR_PH]]:
+; UNROLL-NO-VF-NEXT:    [[N_MOD_VF:%.*]] = urem i64 [[TMP1]], 2
+; UNROLL-NO-VF-NEXT:    [[N_VEC:%.*]] = sub i64 [[TMP1]], [[N_MOD_VF]]
+; UNROLL-NO-VF-NEXT:    br label %[[VECTOR_BODY:.*]]
+; UNROLL-NO-VF:       [[VECTOR_BODY]]:
+; UNROLL-NO-VF-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]
+; UNROLL-NO-VF-NEXT:    [[TMP0:%.*]] = load i32, ptr [[SRC]], align 4
+; UNROLL-NO-VF-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 2
+; UNROLL-NO-VF-NEXT:    [[TMP2:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
+; UNROLL-NO-VF-NEXT:    br i1 [[TMP2]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]
+; UNROLL-NO-VF:       [[MIDDLE_BLOCK]]:
+; UNROLL-NO-VF-NEXT:    [[CMP_N:%.*]] = icmp eq i64 [[TMP1]], [[N_VEC]]
+; UNROLL-NO-VF-NEXT:    br i1 [[CMP_N]], label %[[EXIT:.*]], label %[[SCALAR_PH]]
+; UNROLL-NO-VF:       [[SCALAR_PH]]:
+; UNROLL-NO-VF-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC]], %[[MIDDLE_BLOCK]] ], [ 0, %[[ENTRY]] ]
+; UNROLL-NO-VF-NEXT:    [[SCALAR_RECUR_INIT:%.*]] = phi i32 [ [[TMP0]], %[[MIDDLE_BLOCK]] ], [ 0, %[[ENTRY]] ]
+; UNROLL-NO-VF-NEXT:    br label %[[LOOP:.*]]
+; UNROLL-NO-VF:       [[LOOP]]:
+; UNROLL-NO-VF-NEXT:    [[PHI:%.*]] = phi i64 [ [[BC_RESUME_VAL]], %[[SCALAR_PH]] ], [ [[ADD:%.*]], %[[LOOP]] ]
+; UNROLL-NO-VF-NEXT:    [[RECUR:%.*]] = phi i32 [ [[SCALAR_RECUR_INIT]], %[[SCALAR_PH]] ], [ [[LOAD:%.*]], %[[LOOP]] ]
+; UNROLL-NO-VF-NEXT:    [[ADD]] = add i64 [[PHI]], 1
+; UNROLL-NO-VF-NEXT:    [[LOAD]] = load i32, ptr [[SRC]], align 4
+; UNROLL-NO-VF-NEXT:    [[ICMP:%.*]] = icmp ult i64 [[PHI]], [[N]]
+; UNROLL-NO-VF-NEXT:    br i1 [[ICMP]], label %[[LOOP]], label %[[EXIT]], !llvm.loop [[LOOP3:![0-9]+]]
+; UNROLL-NO-VF:       [[EXIT]]:
+; UNROLL-NO-VF-NEXT:    [[RECUR_LCSSA:%.*]] = phi i32 [ [[RECUR]], %[[LOOP]] ], [ [[TMP0]], %[[MIDDLE_BLOCK]] ]
+; UNROLL-NO-VF-NEXT:    ret i32 [[RECUR_LCSSA]]
+;
+; SINK-AFTER-LABEL: define i32 @recurence_uniform_load(
+; SINK-AFTER-SAME: ptr [[SRC:%.*]], ptr noalias [[DST:%.*]], i64 [[N:%.*]]) {
+; SINK-AFTER-NEXT:  [[ENTRY:.*]]:
+; SINK-AFTER-NEXT:    [[TMP0:%.*]] = add i64 [[N]], 1
+; SINK-AFTER-NEXT:    [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[TMP0]], 4
+; SINK-AFTER-NEXT:    br i1 [[MIN_ITERS_CHECK]], label %[[SCALAR_PH:.*]], label %[[VECTOR_PH:.*]]
+; SINK-AFTER:       [[VECTOR_PH]]:
+; SINK-AFTER-NEXT:    [[N_MOD_VF:%.*]] = urem i64 [[TMP0]], 4
+; SINK-AFTER-NEXT:    [[N_VEC:%.*]] = sub i64 [[TMP0]], [[N_MOD_VF]]
+; SINK-AFTER-NEXT:    br label %[[VECTOR_BODY:.*]]
+; SINK-AFTER:       [[VECTOR_BODY]]:
+; SINK-AFTER-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]
+; SINK-AFTER-NEXT:    [[TMP1:%.*]] = load i32, ptr [[SRC]], align 4
+; SINK-AFTER-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 4
+; SINK-AFTER-NEXT:    [[TMP2:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
+; SINK-AFTER-NEXT:    br i1 [[TMP2]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]
+; SINK-AFTER:       [[MIDDLE_BLOCK]]:
+; SINK-AFTER-NEXT:    [[CMP_N:%.*]] = icmp eq i64 [[TMP0]], [[N_VEC]]
+; SINK-AFTER-NEXT:    br i1 [[CMP_N]], label %[[EXIT:.*]], label %[[SCALAR_PH]]
+; SINK-AFTER:       [[SCALAR_PH]]:
+; SINK-AFTER-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC]], %[[MIDDLE_BLOCK]] ], [ 0, %[[ENTRY]] ]
+; SINK-AFTER-NEXT:    [[SCALAR_RECUR_INIT:%.*]] = phi i32 [ [[TMP1]], %[[MIDDLE_BLOCK]] ], [ 0, %[[ENTRY]] ]
+; SINK-AFTER-NEXT:    br label %[[LOOP:.*]]
+; SINK-AFTER:       [[LOOP]]:
+; SINK-AFTER-NEXT:    [[PHI:%.*]] = phi i64 [ [[BC_RESUME_VAL]], %[[SCALAR_PH]] ], [ [[ADD:%.*]], %[[LOOP]] ]
+; SINK-AFTER-NEXT:    [[RECUR:%.*]] = phi i32 [ [[SCALAR_RECUR_INIT]], %[[SCALAR_PH]] ], [ [[LOAD:%.*]], %[[LOOP]] ]
+; SINK-AFTER-NEXT:    [[ADD]] = add i64 [[PHI]], 1
+; SINK-AFTER-NEXT:    [[LOAD]] = load i32, ptr [[SRC]], align 4
+; SINK-AFTER-NEXT:    [[ICMP:%.*]] = icmp ult i64 [[PHI]], [[N]]
+; SINK-AFTER-NEXT:    br i1 [[ICMP]], label %[[LOOP]], label %[[EXIT]], !llvm.loop [[LOOP3:![0-9]+]]
+; SINK-AFTER:       [[EXIT]]:
+; SINK-AFTER-NEXT:    [[RECUR_LCSSA:%.*]] = phi i32 [ [[RECUR]], %[[LOOP]] ], [ [[TMP1]], %[[MIDDLE_BLOCK]] ]
+; SINK-AFTER-NEXT:    ret i32 [[RECUR_LCSSA]]
+;
+entry:
+  br label %loop
+
+loop:
+  %phi = phi i64 [ 0, %entry ], [ %add, %loop ]
+  %recur = phi i32 [ 0, %entry ], [ %load, %loop ]
+  %add = add i64 %phi, 1
+  %load = load i32, ptr %src, align 4
+  %icmp = icmp ult i64 %phi, %n
+  br i1 %icmp, label %loop, label %exit
+
+exit:
+  ret i32 %recur
+}
+
+; Test for https://github.com/llvm/llvm-project/issues/158319. The recurrence
+; phi can be removed.
+define i16 @for_phi_removed(ptr  %src) {
+; UNROLL-NO-IC-LABEL: define i16 @for_phi_removed(
+; UNROLL-NO-IC-SAME: ptr [[SRC:%.*]]) {
+; UNROLL-NO-IC-NEXT:  [[ENTRY:.*]]:
+; UNROLL-NO-IC-NEXT:    br i1 false, label %[[SCALAR_PH:.*]], label %[[VECTOR_PH:.*]]
+; UNROLL-NO-IC:       [[VECTOR_PH]]:
+; UNROLL-NO-IC-NEXT:    br label %[[VECTOR_BODY:.*]]
+; UNROLL-NO-IC:       [[VECTOR_BODY]]:
+; UNROLL-NO-IC-NEXT:    [[INDEX:%.*]] = phi i32 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]
+; UNROLL-NO-IC-NEXT:    [[TMP0:%.*]] = load i32, ptr [[SRC]], align 4
+; UNROLL-NO-IC-NEXT:    [[BROADCAST_SPLATINSERT:%.*]] = insertelement <4 x i32> poison, i32 [[TMP0]], i64 0
+; UNROLL-NO-IC-NEXT:    [[BROADCAST_SPLAT:%.*]] = shufflevector <4 x i32> [[BROADCAST_SPLATINSERT]], <4 x i32> poison, <4 x i32> zeroinitializer
+; UNROLL-NO-IC-NEXT:    [[TMP1:%.*]] = icmp eq <4 x i32> [[BROADCAST_SPLAT]], zeroinitializer
+; UNROLL-NO-IC-NEXT:    [[TMP2:%.*]] = select <4 x i1> [[TMP1]], <4 x i16> splat (i16 1), <4 x i16> zeroinitializer
+; UNROLL-NO-IC-NEXT:    [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 8
+; UNROLL-NO-IC-NEXT:    [[TMP3:%.*]] = icmp eq i32 [[INDEX_NEXT]], 104
+; UNROLL-NO-IC-NEXT:    br i1 [[TMP3]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP4:![0-9]+]]
+; UNROLL-NO-IC:       [[MIDDLE_BLOCK]]:
+; UNROLL-NO-IC-NEXT:    [[VECTOR_RECUR_EXTRACT:%.*]] = extractelement <4 x i16> [[TMP2]], i32 3
+; UNROLL-NO-IC-NEXT:    br label %[[SCALAR_PH]]
+; UNROLL-NO-IC:       [[SCALAR_PH]]:
+; UNROLL-NO-IC-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i16 [ 104, %[[MIDDLE_BLOCK]] ], [ 0, %[[ENTRY]] ]
+; UNROLL-NO-IC-NEXT:    [[SCALAR_RECUR_INIT:%.*]] = phi i16 [ [[VECTOR_RECUR_EXTRACT]], %[[MIDDLE_BLOCK]] ], [ 0, %[[ENTRY]] ]
+; UNROLL-NO-IC-NEXT:    br label %[[LOOP:.*]]
+; UNROLL-NO-IC:       [[LOOP]]:
+; UNROLL-NO-IC-NEXT:    [[IV:%.*]] = phi i16 [ [[BC_RESUME_VAL]], %[[SCALAR_PH]] ], [ [[IV_NEXT:%.*]], %[[LOOP]] ]
+; UNROLL-NO-IC-NEXT:    [[P:%.*]] = phi i16 [ [[SCALAR_RECUR_INIT]], %[[SCALAR_PH]] ], [ [[SEL:%.*]], %[[LOOP]] ]
+; UNROLL-NO-IC-NEXT:    [[L:%.*]] = load i32, ptr [[SRC]], align 4
+; UNROLL-NO-IC-NEXT:    [[C:%.*]] = icmp eq i32 [[L]], 0
+; UNROLL-NO-IC-NEXT:    [[SEL]] = select i1 [[C]], i16 1, i16 0
+; UNROLL-NO-IC-NEXT:    [[IV_NEXT]] = add i16 [[IV]], 1
+; UNROLL-NO-IC-NEXT:    [[EC:%.*]] = icmp eq i16 [[IV_NEXT]], 111
+; UNROLL-NO-IC-NEXT:    br i1 [[EC]], label %[[EXIT:.*]], label %[[LOOP]], !llvm.loop [[LOOP5:![0-9]+]]
+; UNROLL-NO-IC:       [[EXIT]]:
+; UNROLL-NO-IC-NEXT:    [[P_LCSSA:%.*]] = phi i16 [ [[P]], %[[LOOP]] ]
+; UNROLL-NO-IC-NEXT:    ret i16 [[P_LCSSA]]
+;
+; UNROLL-NO-VF-LABEL: define i16 @for_phi_removed(
+; UNROLL-NO-VF-SAME: ptr [[SRC:%.*]]) {
+; UNROLL-NO-VF-NEXT:  [[ENTRY:.*]]:
+; UNROLL-NO-VF-NEXT:    br i1 false, label %[[SCALAR_PH:.*]], label %[[VECTOR_PH:.*]]
+; UNROLL-NO-VF:       [[VECTOR_PH]]:
+; UNROLL-NO-VF-NEXT:    br label %[[VECTOR_BODY:.*]]
+; UNROLL-NO-VF:       [[VECTOR_BODY]]:
+; UNROLL-NO-VF-NEXT:    [[INDEX:%.*]] = phi i32 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]
+; UNROLL-NO-VF-NEXT:    [[TMP0:%.*]] = load i32, ptr [[SRC]], align 4
+; UNROLL-NO-VF-NEXT:    [[TMP1:%.*]] = icmp eq i32 [[TMP0]], 0
+; UNROLL-NO-VF-NEXT:    [[TMP2:%.*]] = select i1 [[TMP1]], i16 1, i16 0
+; UNROLL-NO-VF-NEXT:    [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 2
+; UNROLL-NO-VF-NEXT:    [[TMP3:%.*]] = icmp eq i32 [[INDEX_NEXT]], 110
+; UNROLL-NO-VF-NEXT:    br i1 [[TMP3]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP4:![0-9]+]]
+; UNROLL-NO-VF:       [[MIDDLE_BLOCK]]:
+; UNROLL-NO-VF-NEXT:    br label %[[SCALAR_PH]]
+; UNROLL-NO-VF:       [[SCALAR_PH]]:
+; UNROLL-NO-VF-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i16 [ 110, %[[MIDDLE_BLOCK]] ], [ 0, %[[ENTRY]] ]
+; UNROLL-NO-VF-NEXT:    [[SCALAR_RECUR_INIT:%.*]] = phi i16 [ [[TMP2]], %[[MIDDLE_BLOCK]] ], [ 0, %[[ENTRY]] ]
+; UNROLL-NO-VF-NEXT:    br label %[[LOOP:.*]]
+; UNROLL-NO-VF:       [[LOOP]]:
+; UNROLL-NO-VF-NEXT:    [[IV:%.*]] = phi i16 [ [[BC_RESUME_VAL]], %[[SCALAR_PH]] ], [ [[IV_NEXT:%.*]], %[[LOOP]] ]
+; UNROLL-NO-VF-NEXT:    [[P:%.*]] = phi i16 [ [[SCALAR_RECUR_INIT]], %[[SCALAR_PH]] ], [ [[SEL:%.*]], %[[LOOP]] ]
+; UNROLL-NO-VF-NEXT:    [[L:%.*]] = load i32, ptr [[SRC]], align 4
+; UNROLL-NO-VF-NEXT:    [[C:%.*]] = icmp eq i32 [[L]], 0
+; UNROLL-NO-VF-NEXT:    [[SEL]] = select i1 [[C]], i16 1, i16 0
+; UNROLL-NO-VF-NEXT:    [[IV_NEXT]] = add i16 [[IV]], 1
+; UNROLL-NO-VF-NEXT:    [[EC:%.*]] = icmp eq i16 [[IV_NEXT]], 111
+; UNROLL-NO-VF-NEXT:    br i1 [[EC]], label %[[EXIT:.*]], label %[[LOOP]], !llvm.loop [[LOOP5:![0-9]+]]
+; UNROLL-NO-VF:       [[EXIT]]:
+; UNROLL-NO-VF-NEXT:    [[P_LCSSA:%.*]] = phi i16 [ [[P]], %[[LOOP]] ]
+; UNROLL-NO-VF-NEXT:    ret i16 [[P_LCSSA]]
+;
+; SINK-AFTER-LABEL: define i16 @for_phi_removed(
+; SINK-AFTER-SAME: ptr [[SRC:%.*]]) {
+; SINK-AFTER-NEXT:  [[ENTRY:.*]]:
+; SINK-AFTER-NEXT:    br i1 false, label %[[SCALAR_PH:.*]], label %[[VECTOR_PH:.*]]
+; SINK-AFTER:       [[VECTOR_PH]]:
+; SINK-AFTER-NEXT:    br label %[[VECTOR_BODY:.*]]
+; SINK-AFTER:       [[VECTOR_BODY]]:
+; SINK-AFTER-NEXT:    [[INDEX:%.*]] = phi i32 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]
+; SINK-AFTER-NEXT:    [[TMP0:%.*]] = load i32, ptr [[SRC]], align 4
+; SINK-AFTER-NEXT:    [[BROADCAST_SPLATINSERT:%.*]] = insertelement <4 x i32> poison, i32 [[TMP0]], i64 0
+; SINK-AFTER-NEXT:    [[BROADCAST_SPLAT:%.*]] = shufflevector <4 x i32> [[BROADCAST_SPLATINSERT]], <4 x i32> poison, <4 x i32> zeroinitializer
+; SINK-AFTER-NEXT:    [[TMP1:%.*]] = icmp eq <4 x i32> [[BROADCAST_SPLAT]], zeroinitializer
+; SINK-AFTER-NEXT:    [[TMP2:%.*]] = select <4 x i1> [[TMP1]], <4 x i16> splat (i16 1), <4 x i16> zeroinitializer
+; SINK-AFTER-NEXT:    [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 4
+; SINK-AFTER-NEXT:    [[TMP3:%.*]] = icmp eq i32 [[INDEX_NEXT]], 108
+; SINK-AFTER-NEXT:    br i1 [[TMP3]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP4:![0-9]+]]
+; SINK-AFTER:       [[MIDDLE_BLOCK]]:
+; SINK-AFTER-NEXT:    [[VECTOR_RECUR_EXTRACT:%.*]] = extractelement <4 x i16> [[TMP2]], i32 3
+; SINK-AFTER-NEXT:    br label %[[SCALAR_PH]]
+; SINK-AFTER:       [[SCALAR_PH]]:
+; SINK-AFTER-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i16 [ 108, %[[MIDDLE_BLOCK]] ], [ 0, %[[ENTRY]] ]
+; SINK-AFTER-NEXT:    [[SCALAR_RECUR_INIT:%.*]] = phi i16 [ [[VECTOR_RECUR_EXTRACT]], %[[MIDDLE_BLOCK]] ], [ 0, %[[ENTRY]] ]
+; SINK-AFTER-NEXT:    br label %[[LOOP:.*]]
+; SINK-AFTER:       [[LOOP]]:
+; SINK-AFTER-NEXT:    [[IV:%.*]] = phi i16 [ [[BC_RESUME_VAL]], %[[SCALAR_PH]] ], [ [[IV_NEXT:%.*]], %[[LOOP]] ]
+; SINK-AFTER-NEXT:    [[P:%.*]] = phi i16 [ [[SCALAR_RECUR_INIT]], %[[SCALAR_PH]] ], [ [[SEL:%.*]], %[[LOOP]] ]
+; SINK-AFTER-NEXT:    [[L:%.*]] = load i32, ptr [[SRC]], align 4
+; SINK-AFTER-NEXT:    [[C:%.*]] = icmp eq i32 [[L]], 0
+; SINK-AFTER-NEXT:    [[SEL]] = select i1 [[C]], i16 1, i16 0
+; SINK-AFTER-NEXT:    [[IV_NEXT]] = add i16 [[IV]], 1
+; SINK-AFTER-NEXT:    [[EC:%.*]] = icmp eq i16 [[IV_NEXT]], 111
+; SINK-AFTER-NEXT:    br i1 [[EC]], label %[[EXIT:.*]], label %[[LOOP]], !llvm.loop [[LOOP5:![0-9]+]]
+; SINK-AFTER:       [[EXIT]]:
+; SINK-AFTER-NEXT:    [[P_LCSSA:%.*]] = phi i16 [ [[P]], %[[LOOP]] ]
+; SINK-AFTER-NEXT:    ret i16 [[P_LCSSA]]
+;
+entry:
+  br label %loop
+
+loop:
+  %iv = phi i16 [ 0, %entry ], [ %iv.next, %loop ]
+  %p = phi i16 [ 0, %entry ] , [ %sel, %loop ]
+  %l = load i32, ptr %src, align 4
+  %c = icmp eq i32 %l, 0
+  %sel = select i1 %c, i16 1, i16 0
+  %iv.next = add i16 %iv, 1
+  %ec = icmp eq i16 %iv.next, 111
+  br i1 %ec, label %exit, label %loop
+
+exit:
+  ret i16 %p
+}