[LV] Use ExtractLane(LastActiveLane, V) live outs when tail-folding.

llvm/lib/Transforms/Vectorize/LoopVectorizationLegality.cpp

@@ -2083,24 +2083,6 @@ bool LoopVectorizationLegality::canFoldTailByMasking() const {
   for (const auto &Reduction : getReductionVars())
     ReductionLiveOuts.insert(Reduction.second.getLoopExitInstr());
 
-  // TODO: handle non-reduction outside users when tail is folded by masking.
-  for (auto *AE : AllowedExit) {
-    // Check that all users of allowed exit values are inside the loop or
-    // are the live-out of a reduction.
-    if (ReductionLiveOuts.count(AE))
-      continue;
-    for (User *U : AE->users()) {
-      Instruction *UI = cast<Instruction>(U);
-      if (TheLoop->contains(UI))
-        continue;
-      LLVM_DEBUG(
-          dbgs()
-          << "LV: Cannot fold tail by masking, loop has an outside user for "
-          << *UI << "\n");
-      return false;
-    }
-  }
-
   for (const auto &Entry : getInductionVars()) {
     PHINode *OrigPhi = Entry.first;
     for (User *U : OrigPhi->users()) {

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

@@ -8798,7 +8798,11 @@ void LoopVectorizationPlanner::adjustRecipesForReductions(
       if (FinalReductionResult == U || Parent->getParent())
         continue;
       U->replaceUsesOfWith(OrigExitingVPV, FinalReductionResult);
-      if (match(U, m_ExtractLastElement(m_VPValue())))
+      if (match(U,
+                m_CombineOr(m_VPInstruction<VPInstruction::ExtractLastElement>(
+                                m_VPValue()),
+                            m_VPInstruction<VPInstruction::ExtractLane>(
+                                m_VPValue(), m_VPValue()))))
         cast<VPInstruction>(U)->replaceAllUsesWith(FinalReductionResult);
     }
 

llvm/lib/Transforms/Vectorize/VPlanPredicator.cpp

@@ -44,11 +44,6 @@ class VPPredicator {
   /// possibly inserting new recipes at \p Dst (using Builder's insertion point)
   VPValue *createEdgeMask(VPBasicBlock *Src, VPBasicBlock *Dst);
 
-  /// Returns the *entry* mask for \p VPBB.
-  VPValue *getBlockInMask(VPBasicBlock *VPBB) const {
-    return BlockMaskCache.lookup(VPBB);
-  }
-
   /// Record \p Mask as the *entry* mask of \p VPBB, which is expected to not
   /// already have a mask.
   void setBlockInMask(VPBasicBlock *VPBB, VPValue *Mask) {
@@ -68,6 +63,11 @@ class VPPredicator {
   }
 
 public:
+  /// Returns the *entry* mask for \p VPBB.
+  VPValue *getBlockInMask(VPBasicBlock *VPBB) const {
+    return BlockMaskCache.lookup(VPBB);
+  }
+
   /// Returns the precomputed predicate of the edge from \p Src to \p Dst.
   VPValue *getEdgeMask(const VPBasicBlock *Src, const VPBasicBlock *Dst) const {
     return EdgeMaskCache.lookup({Src, Dst});
@@ -300,5 +300,46 @@ VPlanTransforms::introduceMasksAndLinearize(VPlan &Plan, bool FoldTail) {
 
     PrevVPBB = VPBB;
   }
+
+  // If we folded the tail and introduced a header mask, any extract of the
+  // last element must be updated to extract from the last active lane of the
+  // header mask instead (i.e., the lane corresponding to the last active
+  // iteration).
+  if (FoldTail) {
+    assert(Plan.getExitBlocks().size() == 1 &&
+           "only a single-exit block is supported currently");
+    VPBasicBlock *EB = Plan.getExitBlocks().front();
+    assert(EB->getSinglePredecessor() == Plan.getMiddleBlock() &&
+           "the exit block must have middle block as single predecessor");
+
+    VPValue *LastActiveLane = nullptr;
+    VPBuilder B(Plan.getMiddleBlock()->getTerminator());
+    for (auto &P : EB->phis()) {
+      auto *ExitIRI = cast<VPIRPhi>(&P);
+      VPValue *Inc = ExitIRI->getIncomingValue(0);
+      VPValue *Op;
+      if (!match(Inc, m_VPInstruction<VPInstruction::ExtractLastElement>(
+                          m_VPValue(Op))))
+        continue;
+
+      if (!LastActiveLane) {
+        // Compute the index of the last active lane by getting the first lane
+        // where the header mask is false (first inactive lane), then
+        // subtracting 1. This gives us the last lane where the mask was true.
+        VPValue *HeaderMask = Predicator.getBlockInMask(
+            Plan.getVectorLoopRegion()->getEntryBasicBlock());
+        VPValue *NotHeaderMask = B.createNot(HeaderMask);
+        VPValue *FirstInactiveLane =
+            B.createNaryOp(VPInstruction::FirstActiveLane, {NotHeaderMask});
+        VPValue *One = Plan.getOrAddLiveIn(
+            ConstantInt::get(Type::getInt64Ty(Plan.getContext()), 1));
+        LastActiveLane =
+            B.createNaryOp(Instruction::Sub, {FirstInactiveLane, One});
+      }
+      auto *Ext =
+          B.createNaryOp(VPInstruction::ExtractLane, {LastActiveLane, Op});
+      Inc->replaceAllUsesWith(Ext);
+    }
+  }
   return Predicator.getBlockMaskCache();
 }

llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp

@@ -638,7 +638,12 @@ Value *VPInstruction::generate(VPTransformState &State) {
     llvm_unreachable("should be handled by VPPhi::execute");
   }
   case Instruction::Select: {
-    bool OnlyFirstLaneUsed = vputils::onlyFirstLaneUsed(this);
+    bool OnlyFirstLaneUsed =
+        vputils::onlyFirstLaneUsed(this) ||
+        (isa<VPInstruction>(getOperand(1)) &&
+         cast<VPInstruction>(getOperand(1))->isVectorToScalar() &&
+         isa<VPInstruction>(getOperand(2)) &&
+         cast<VPInstruction>(getOperand(2))->isVectorToScalar());
     Value *Cond = State.get(getOperand(0), OnlyFirstLaneUsed);
     Value *Op1 = State.get(getOperand(1), OnlyFirstLaneUsed);
     Value *Op2 = State.get(getOperand(2), OnlyFirstLaneUsed);

llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp

@@ -1941,6 +1941,35 @@ bool VPlanTransforms::adjustFixedOrderRecurrences(VPlan &Plan,
     // Set the first operand of RecurSplice to FOR again, after replacing
     // all users.
     RecurSplice->setOperand(0, FOR);
+
+    // Check for users extracting the second-to-last active lane of the FOR. If
+    // only a single lane is active in the current iteration, we need to select
+    // the last element of the value from the previous iteration, directly from
+    // the FOR phi.
+    for (VPUser *U : RecurSplice->users()) {
+      if (!match(U, m_VPInstruction<VPInstruction::ExtractLane>(
+                        m_Sub(m_VPInstruction<VPInstruction::FirstActiveLane>(
+                                  m_VPValue()),
+                              m_SpecificInt(1)),
+                        m_Specific(RecurSplice))))
+        continue;
+
+      VPBuilder B(cast<VPInstruction>(U));
+      VPValue *LastActiveLane = cast<VPInstruction>(U)->getOperand(0);
+      Type *I64Ty = Type::getInt64Ty(Plan.getContext());
+      VPValue *Zero = Plan.getOrAddLiveIn(ConstantInt::get(I64Ty, 0));
+      VPValue *One = Plan.getOrAddLiveIn(ConstantInt::get(I64Ty, 1));
+      VPValue *PenultimateIndex =
+          B.createNaryOp(Instruction::Sub, {LastActiveLane, One});
+      VPValue *PenultimateLastIter =
+          B.createNaryOp(VPInstruction::ExtractLane,
+                         {PenultimateIndex, FOR->getBackedgeValue()});
+      VPValue *LastPrevIter =
+          B.createNaryOp(VPInstruction::ExtractLastElement, {FOR});
+      VPValue *Cmp = B.createICmp(CmpInst::ICMP_EQ, LastActiveLane, Zero);
+      VPValue *Sel = B.createSelect(Cmp, LastPrevIter, PenultimateLastIter);
+      cast<VPInstruction>(U)->replaceAllUsesWith(Sel);
+    }
   }
   return true;
 }

llvm/lib/Transforms/Vectorize/VPlanUnroll.cpp

@@ -377,6 +377,12 @@ void UnrollState::unrollBlock(VPBlockBase *VPB) {
         match(&R, m_VPInstruction<VPInstruction::ExtractPenultimateElement>(
                       m_VPValue(Op0)))) {
       addUniformForAllParts(cast<VPSingleDefRecipe>(&R));
+      if (isa<VPFirstOrderRecurrencePHIRecipe>(Op0)) {
+        assert(match(&R, m_ExtractLastElement(m_VPValue())) &&
+               "can only extract last element of FOR");
+        continue;
+      }
+
       if (Plan.hasScalarVFOnly()) {
         auto *I = cast<VPInstruction>(&R);
         // Extracting from end with VF = 1 implies retrieving the last or

llvm/test/Transforms/LoopVectorize/RISCV/dead-ops-cost.ll

@@ -69,61 +69,61 @@ exit:
 define i8 @dead_live_out_due_to_scalar_epilogue_required(ptr %src, ptr %dst) {
 ; CHECK-LABEL: define i8 @dead_live_out_due_to_scalar_epilogue_required(
 ; CHECK-SAME: ptr [[SRC:%.*]], ptr [[DST:%.*]]) #[[ATTR0]] {
-; CHECK-NEXT:  [[ENTRY:.*]]:
-; CHECK-NEXT:    [[TMP0:%.*]] = call i32 @llvm.vscale.i32()
-; CHECK-NEXT:    [[TMP1:%.*]] = shl nuw i32 [[TMP0]], 2
-; CHECK-NEXT:    [[TMP2:%.*]] = call i32 @llvm.umax.i32(i32 [[TMP1]], i32 6)
-; CHECK-NEXT:    [[MIN_ITERS_CHECK:%.*]] = icmp ule i32 252, [[TMP2]]
-; CHECK-NEXT:    br i1 [[MIN_ITERS_CHECK]], label %[[SCALAR_PH:.*]], label %[[VECTOR_MEMCHECK:.*]]
+; CHECK-NEXT:  [[ENTRY:.*:]]
+; CHECK-NEXT:    br label %[[VECTOR_MEMCHECK:.*]]
 ; CHECK:       [[VECTOR_MEMCHECK]]:
 ; CHECK-NEXT:    [[SCEVGEP:%.*]] = getelementptr i8, ptr [[DST]], i64 1005
 ; CHECK-NEXT:    [[SCEVGEP1:%.*]] = getelementptr i8, ptr [[SRC]], i64 1005
 ; CHECK-NEXT:    [[BOUND0:%.*]] = icmp ult ptr [[DST]], [[SCEVGEP1]]
 ; CHECK-NEXT:    [[BOUND1:%.*]] = icmp ult ptr [[SRC]], [[SCEVGEP]]
 ; CHECK-NEXT:    [[FOUND_CONFLICT:%.*]] = and i1 [[BOUND0]], [[BOUND1]]
-; CHECK-NEXT:    br i1 [[FOUND_CONFLICT]], label %[[SCALAR_PH]], label %[[VECTOR_PH:.*]]
+; CHECK-NEXT:    br i1 [[FOUND_CONFLICT]], label %[[SCALAR_PH:.*]], label %[[VECTOR_PH:.*]]
 ; CHECK:       [[VECTOR_PH]]:
-; CHECK-NEXT:    [[TMP3:%.*]] = call i32 @llvm.vscale.i32()
-; CHECK-NEXT:    [[TMP4:%.*]] = mul nuw i32 [[TMP3]], 4
-; CHECK-NEXT:    [[N_MOD_VF:%.*]] = urem i32 252, [[TMP4]]
-; CHECK-NEXT:    [[TMP5:%.*]] = icmp eq i32 [[N_MOD_VF]], 0
-; CHECK-NEXT:    [[TMP6:%.*]] = select i1 [[TMP5]], i32 [[TMP4]], i32 [[N_MOD_VF]]
-; CHECK-NEXT:    [[N_VEC:%.*]] = sub i32 252, [[TMP6]]
-; CHECK-NEXT:    [[IND_END:%.*]] = mul i32 [[N_VEC]], 4
-; CHECK-NEXT:    [[TMP9:%.*]] = call <vscale x 4 x i32> @llvm.stepvector.nxv4i32()
-; CHECK-NEXT:    [[TMP11:%.*]] = mul <vscale x 4 x i32> [[TMP9]], splat (i32 4)
-; CHECK-NEXT:    [[INDUCTION:%.*]] = add <vscale x 4 x i32> zeroinitializer, [[TMP11]]
-; CHECK-NEXT:    [[TMP14:%.*]] = mul i32 4, [[TMP4]]
-; CHECK-NEXT:    [[DOTSPLATINSERT:%.*]] = insertelement <vscale x 4 x i32> poison, i32 [[TMP14]], i64 0
-; CHECK-NEXT:    [[DOTSPLAT:%.*]] = shufflevector <vscale x 4 x i32> [[DOTSPLATINSERT]], <vscale x 4 x i32> poison, <vscale x 4 x i32> zeroinitializer
+; CHECK-NEXT:    [[TMP0:%.*]] = call <vscale x 16 x i32> @llvm.stepvector.nxv16i32()
+; CHECK-NEXT:    [[TMP1:%.*]] = mul <vscale x 16 x i32> [[TMP0]], splat (i32 4)
+; CHECK-NEXT:    [[INDUCTION:%.*]] = add <vscale x 16 x i32> zeroinitializer, [[TMP1]]
 ; CHECK-NEXT:    br label %[[VECTOR_BODY:.*]]
 ; CHECK:       [[VECTOR_BODY]]:
-; CHECK-NEXT:    [[INDEX:%.*]] = phi i32 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]
-; CHECK-NEXT:    [[VEC_IND:%.*]] = phi <vscale x 4 x i32> [ [[INDUCTION]], %[[VECTOR_PH]] ], [ [[VEC_IND_NEXT:%.*]], %[[VECTOR_BODY]] ]
-; CHECK-NEXT:    [[TMP15:%.*]] = sext <vscale x 4 x i32> [[VEC_IND]] to <vscale x 4 x i64>
-; CHECK-NEXT:    [[TMP16:%.*]] = getelementptr i8, ptr [[DST]], <vscale x 4 x i64> [[TMP15]]
-; CHECK-NEXT:    call void @llvm.masked.scatter.nxv4i8.nxv4p0(<vscale x 4 x i8> zeroinitializer, <vscale x 4 x ptr> [[TMP16]], i32 1, <vscale x 4 x i1> splat (i1 true)), !alias.scope [[META3:![0-9]+]], !noalias [[META6:![0-9]+]]
-; CHECK-NEXT:    [[INDEX_NEXT]] = add nuw i32 [[INDEX]], [[TMP4]]
-; CHECK-NEXT:    [[VEC_IND_NEXT]] = add <vscale x 4 x i32> [[VEC_IND]], [[DOTSPLAT]]
-; CHECK-NEXT:    [[TMP17:%.*]] = icmp eq i32 [[INDEX_NEXT]], [[N_VEC]]
-; CHECK-NEXT:    br i1 [[TMP17]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP8:![0-9]+]]
+; CHECK-NEXT:    [[VEC_IND:%.*]] = phi <vscale x 16 x i32> [ [[INDUCTION]], %[[VECTOR_PH]] ], [ [[VEC_IND_NEXT:%.*]], %[[VECTOR_BODY]] ]
+; CHECK-NEXT:    [[AVL:%.*]] = phi i32 [ 252, %[[VECTOR_PH]] ], [ [[AVL_NEXT:%.*]], %[[VECTOR_BODY]] ]
+; CHECK-NEXT:    [[TMP2:%.*]] = call i32 @llvm.experimental.get.vector.length.i32(i32 [[AVL]], i32 16, i1 true)
+; CHECK-NEXT:    [[BROADCAST_SPLATINSERT2:%.*]] = insertelement <vscale x 16 x i32> poison, i32 [[TMP2]], i64 0
+; CHECK-NEXT:    [[BROADCAST_SPLAT3:%.*]] = shufflevector <vscale x 16 x i32> [[BROADCAST_SPLATINSERT2]], <vscale x 16 x i32> poison, <vscale x 16 x i32> zeroinitializer
+; CHECK-NEXT:    [[TMP3:%.*]] = mul i32 4, [[TMP2]]
+; CHECK-NEXT:    [[BROADCAST_SPLATINSERT:%.*]] = insertelement <vscale x 16 x i32> poison, i32 [[TMP3]], i64 0
+; CHECK-NEXT:    [[BROADCAST_SPLAT:%.*]] = shufflevector <vscale x 16 x i32> [[BROADCAST_SPLATINSERT]], <vscale x 16 x i32> poison, <vscale x 16 x i32> zeroinitializer
+; CHECK-NEXT:    [[TMP5:%.*]] = icmp uge <vscale x 16 x i32> [[TMP0]], [[BROADCAST_SPLAT3]]
+; CHECK-NEXT:    [[TMP9:%.*]] = sext <vscale x 16 x i32> [[VEC_IND]] to <vscale x 16 x i64>
+; CHECK-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[SRC]], <vscale x 16 x i64> [[TMP9]]
+; CHECK-NEXT:    [[WIDE_MASKED_GATHER:%.*]] = call <vscale x 16 x i8> @llvm.vp.gather.nxv16i8.nxv16p0(<vscale x 16 x ptr> align 1 [[TMP6]], <vscale x 16 x i1> splat (i1 true), i32 [[TMP2]]), !alias.scope [[META3:![0-9]+]]
+; CHECK-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[DST]], <vscale x 16 x i64> [[TMP9]]
+; CHECK-NEXT:    call void @llvm.vp.scatter.nxv16i8.nxv16p0(<vscale x 16 x i8> zeroinitializer, <vscale x 16 x ptr> align 1 [[TMP7]], <vscale x 16 x i1> splat (i1 true), i32 [[TMP2]]), !alias.scope [[META6:![0-9]+]], !noalias [[META3]]
+; CHECK-NEXT:    [[AVL_NEXT]] = sub nuw i32 [[AVL]], [[TMP2]]
+; CHECK-NEXT:    [[VEC_IND_NEXT]] = add <vscale x 16 x i32> [[VEC_IND]], [[BROADCAST_SPLAT]]
+; CHECK-NEXT:    [[TMP8:%.*]] = icmp eq i32 [[AVL_NEXT]], 0
+; CHECK-NEXT:    br i1 [[TMP8]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP8:![0-9]+]]
 ; CHECK:       [[MIDDLE_BLOCK]]:
-; CHECK-NEXT:    br label %[[SCALAR_PH]]
+; CHECK-NEXT:    [[TMP10:%.*]] = call i64 @llvm.experimental.cttz.elts.i64.nxv16i1(<vscale x 16 x i1> [[TMP5]], i1 true)
+; CHECK-NEXT:    [[TMP11:%.*]] = sub i64 [[TMP10]], 1
+; CHECK-NEXT:    [[TMP12:%.*]] = call i64 @llvm.vscale.i64()
+; CHECK-NEXT:    [[TMP13:%.*]] = mul nuw i64 [[TMP12]], 16
+; CHECK-NEXT:    [[TMP17:%.*]] = mul i64 [[TMP13]], 0
+; CHECK-NEXT:    [[TMP15:%.*]] = extractelement <vscale x 16 x i8> [[WIDE_MASKED_GATHER]], i64 [[TMP11]]
+; CHECK-NEXT:    br label %[[EXIT:.*]]
 ; CHECK:       [[SCALAR_PH]]:
-; CHECK-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i32 [ [[IND_END]], %[[MIDDLE_BLOCK]] ], [ 0, %[[ENTRY]] ], [ 0, %[[VECTOR_MEMCHECK]] ]
 ; CHECK-NEXT:    br label %[[LOOP:.*]]
 ; CHECK:       [[LOOP]]:
-; CHECK-NEXT:    [[IV:%.*]] = phi i32 [ [[BC_RESUME_VAL]], %[[SCALAR_PH]] ], [ [[IV_NEXT:%.*]], %[[LOOP]] ]
+; CHECK-NEXT:    [[IV:%.*]] = phi i32 [ 0, %[[SCALAR_PH]] ], [ [[IV_NEXT:%.*]], %[[LOOP]] ]
 ; CHECK-NEXT:    [[IDXPROM:%.*]] = sext i32 [[IV]] to i64
 ; CHECK-NEXT:    [[GEP_SRC:%.*]] = getelementptr i8, ptr [[SRC]], i64 [[IDXPROM]]
 ; CHECK-NEXT:    [[L:%.*]] = load i8, ptr [[GEP_SRC]], align 1
 ; CHECK-NEXT:    [[GEP_DST:%.*]] = getelementptr i8, ptr [[DST]], i64 [[IDXPROM]]
 ; CHECK-NEXT:    store i8 0, ptr [[GEP_DST]], align 1
 ; CHECK-NEXT:    [[IV_NEXT]] = add i32 [[IV]], 4
 ; CHECK-NEXT:    [[CMP:%.*]] = icmp ult i32 [[IV]], 1001
-; CHECK-NEXT:    br i1 [[CMP]], label %[[LOOP]], label %[[EXIT:.*]], !llvm.loop [[LOOP9:![0-9]+]]
+; CHECK-NEXT:    br i1 [[CMP]], label %[[LOOP]], label %[[EXIT]], !llvm.loop [[LOOP9:![0-9]+]]
 ; CHECK:       [[EXIT]]:
-; CHECK-NEXT:    [[R:%.*]] = phi i8 [ [[L]], %[[LOOP]] ]
+; CHECK-NEXT:    [[R:%.*]] = phi i8 [ [[L]], %[[LOOP]] ], [ [[TMP15]], %[[MIDDLE_BLOCK]] ]
 ; CHECK-NEXT:    ret i8 [[R]]
 ;
 entry: