SelectionDAG: Teach the legalizer to split SETCC if VSELECT needs splitting too.

The Type Legalizer recognizes that VSELECT needs to be split, because the type
is to wide for the given target. The same does not always apply to SETCC,
because less space is required to encode the result of a comparison. As a result
VSELECT is split and SETCC is unrolled into scalar comparisons.

This commit fixes the issue by checking for VSELECT-SETCC patterns in the DAG
Combiner. If a matching pattern is found, then the result mask of SETCC is
promoted to the expected vector mask for the given target. This mask has usually
te same size as the VSELECT return type (except for Intel KNL). Now the type
legalizer will split both VSELECT and SETCC.

This allows the following X86 DAG Combine code to sucessfully detect the MIN/MAX
pattern. This fixes PR16695, PR17002, and <rdar://problem/14594431>.

llvm-svn: 191130

llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp

@@ -4327,6 +4327,27 @@ SDValue DAGCombiner::visitVSELECT(SDNode *N) {
     }
   }
 
+  // Treat SETCC as a mask and promote the result type based on the targets
+  // expected SETCC result type. This will ensure that SETCC and VSELECT are
+  // both split by the type legalizer. This is done to prevent the type
+  // legalizer from unrolling SETCC into scalar comparions.
+  EVT SelectVT = N->getValueType(0);
+  if (N0.getOpcode() == ISD::SETCC &&
+      N0.getValueType() != getSetCCResultType(SelectVT)) {
+    SDLoc MaskDL(N0);
+    EVT MaskVT = getSetCCResultType(SelectVT);
+
+    SDValue Mask = DAG.getNode(ISD::SETCC, MaskDL, MaskVT, N0->getOperand(0),
+                               N0->getOperand(1), N0->getOperand(2));
+
+    AddToWorkList(Mask.getNode());
+
+    SDValue LHS = N->getOperand(1);
+    SDValue RHS = N->getOperand(2);
+
+    return DAG.getNode(ISD::VSELECT, DL, SelectVT, Mask, LHS, RHS);
+  }
+
   return SDValue();
 }
 

llvm/lib/CodeGen/SelectionDAG/LegalizeTypesGeneric.cpp

@@ -489,6 +489,11 @@ void DAGTypeLegalizer::SplitRes_SELECT(SDNode *N, SDValue &Lo,
   SDValue Cond = N->getOperand(0);
   CL = CH = Cond;
   if (Cond.getValueType().isVector()) {
+    if (Cond.getOpcode() == ISD::SETCC) {
+      assert(Cond.getValueType() == getSetCCResultType(N->getValueType(0)) &&
+             "Condition has not been prepared for split!");
+      GetSplitVector(Cond, CL, CH);
+    } else {
       assert(Cond.getValueType().getVectorElementType() == MVT::i1 &&
              "Condition legalized before result?");
       unsigned NumElements = Cond.getValueType().getVectorNumElements();
@@ -498,6 +503,7 @@ void DAGTypeLegalizer::SplitRes_SELECT(SDNode *N, SDValue &Lo,
       CH = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VCondTy, Cond,
                        DAG.getConstant(NumElements / 2, TLI.getVectorIdxTy()));
     }
+  }
 
   Lo = DAG.getNode(N->getOpcode(), dl, LL.getValueType(), CL, LL, RL);
   Hi = DAG.getNode(N->getOpcode(), dl, LH.getValueType(), CH, LH, RH);

llvm/lib/Target/X86/X86ISelLowering.cpp

@@ -1535,7 +1535,16 @@ void X86TargetLowering::resetOperationActions() {
 }
 
 EVT X86TargetLowering::getSetCCResultType(LLVMContext &, EVT VT) const {
-  if (!VT.isVector()) return MVT::i8;
+  if (!VT.isVector())
+    return MVT::i8;
+
+  const TargetMachine &TM = getTargetMachine();
+  if (!TM.Options.UseSoftFloat && Subtarget->hasAVX512())
+    switch(VT.getVectorNumElements()) {
+    case  8: return MVT::v8i1;
+    case 16: return MVT::v16i1;
+    }
+
   return VT.changeVectorElementTypeToInteger();
 }
 

llvm/test/CodeGen/X86/vec_split.ll

@@ -0,0 +1,29 @@
+; RUN: llc -march=x86-64 -mcpu=corei7 < %s | FileCheck %s -check-prefix=SSE4
+; RUN: llc -march=x86-64 -mcpu=core-avx2 < %s | FileCheck %s -check-prefix=AVX2
+
+define <16 x i16> @split16(<16 x i16> %a, <16 x i16> %b, <16 x i8> %__mask) {
+; SSE4-LABEL: split16:
+; SSE4: pminuw
+; SSE4: pminuw
+; AVX2-LABEL: split16:
+; AVX2: vpminuw
+; AVX2: ret
+  %1 = icmp ult <16 x i16> %a, %b
+  %2 = select <16 x i1> %1, <16 x i16> %a, <16 x i16> %b
+  ret <16 x i16> %2
+}
+
+define <32 x i16> @split32(<32 x i16> %a, <32 x i16> %b, <32 x i8> %__mask) {
+; SSE4-LABEL: split32:
+; SSE4: pminuw
+; SSE4: pminuw
+; SSE4: pminuw
+; SSE4: pminuw
+; AVX2-LABEL: split32:
+; AVX2: vpminuw
+; AVX2: vpminuw
+; AVX2: ret
+  %1 = icmp ult <32 x i16> %a, %b
+  %2 = select <32 x i1> %1, <32 x i16> %a, <32 x i16> %b
+  ret <32 x i16> %2
+}