[X86][AVX512] Adding a pattern for broadcastm intrinsic.

Differential Revision: https://reviews.llvm.org/D38312

Change-Id: I6551fb13879e098aed74de410e29815cf37d9ab5
llvm-svn: 316890

llvm/lib/Target/X86/X86ISelLowering.cpp

@@ -6687,6 +6687,44 @@ static bool isUseOfShuffle(SDNode *N) {
   return false;
 }
 
+// Check if the current node of build vector is a zero extended vector.
+// If so, return the value extended.
+// For example: (0,0,0,a,0,0,0,a,0,0,0,a,0,0,0,a) returns a.
+// NumElt - return the number of zero extended identical values.
+// EltType - return the type of the value include the zero extend.
+static SDValue isSplatZeroExtended(const BuildVectorSDNode *Op,
+                                   unsigned &NumElt, MVT &EltType) {
+  SDValue ExtValue = Op->getOperand(0);
+  unsigned NumElts = Op->getNumOperands();
+  unsigned Delta = NumElts;
+
+  for (unsigned i = 1; i < NumElts; i++) {
+    if (Op->getOperand(i) == ExtValue) {
+      Delta = i;
+      break;
+    }
+    if (!(Op->getOperand(i).isUndef() || isNullConstant(Op->getOperand(i))))
+      return SDValue();
+  }
+  if (!isPowerOf2_32(Delta) || Delta == 1)
+    return SDValue();
+
+  for (unsigned i = Delta; i < NumElts; i++) {
+    if (i % Delta == 0) {
+      if (Op->getOperand(i) != ExtValue)
+        return SDValue();
+    } else if (!(isNullConstant(Op->getOperand(i)) ||
+                 Op->getOperand(i).isUndef()))
+      return SDValue();
+  }
+  unsigned EltSize =
+      Op->getSimpleValueType(0).getScalarSizeInBits();
+  unsigned ExtVTSize = EltSize * Delta;
+  EltType = MVT::getIntegerVT(ExtVTSize);
+  NumElt = NumElts / Delta;
+  return ExtValue;
+}
+
 /// Attempt to use the vbroadcast instruction to generate a splat value
 /// from a splat BUILD_VECTOR which uses:
 ///  a. A single scalar load, or a constant.
@@ -6709,6 +6747,32 @@ static SDValue lowerBuildVectorAsBroadcast(BuildVectorSDNode *BVOp,
   assert((VT.is128BitVector() || VT.is256BitVector() || VT.is512BitVector()) &&
          "Unsupported vector type for broadcast.");
 
+  // Attempt to use VBROADCASTM
+  // From this paterrn:
+  //  a. t0 = (zext_i64 (bitcast_i8 v2i1 X))
+  //  b. t1 = (build_vector t0 t0)
+  //
+  // Create (VBROADCASTM v2i1 X)
+  if (Subtarget.hasCDI() && (VT.is512BitVector() || Subtarget.hasVLX())) {
+    MVT EltType;
+    unsigned NumElts;
+    SDValue ZeroExtended = isSplatZeroExtended(BVOp, NumElts, EltType);
+    if (ZeroExtended && ZeroExtended.getOpcode() == ISD::BITCAST) {
+      SDValue BOperand = ZeroExtended.getOperand(0);
+      if (BOperand.getSimpleValueType().getVectorElementType() == MVT::i1) {
+        if ((EltType == MVT::i64 &&
+             VT.getVectorElementType() == MVT::i8) || // for broadcastmb2q
+            (EltType == MVT::i32 &&
+             VT.getVectorElementType() == MVT::i16)) { // for broadcastmw2d
+          SDValue Brdcst =
+              DAG.getNode(X86ISD::VBROADCASTM, dl,
+                          MVT::getVectorVT(EltType, NumElts), BOperand);
+          return DAG.getBitcast(VT, Brdcst);
+        }
+      }
+    }
+  }
+
   BitVector UndefElements;
   SDValue Ld = BVOp->getSplatValue(&UndefElements);
 

llvm/test/CodeGen/X86/broadcastm-lowering.ll

@@ -20,10 +20,7 @@ define <2 x i64> @test_mm_epi64(<8 x i16> %a, <8 x i16> %b) {
 ; AVX512VLCDBW-LABEL: test_mm_epi64:
 ; AVX512VLCDBW:       # BB#0: # %entry
 ; AVX512VLCDBW-NEXT:    vpcmpeqw %xmm1, %xmm0, %k0
-; AVX512VLCDBW-NEXT:    kmovd %k0, %eax
+; AVX512VLCDBW-NEXT:    vpbroadcastmb2q %k0, %xmm0
-; AVX512VLCDBW-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; AVX512VLCDBW-NEXT:    vpinsrb $0, %eax, %xmm0, %xmm0
-; AVX512VLCDBW-NEXT:    vpinsrb $8, %eax, %xmm0, %xmm0
 ; AVX512VLCDBW-NEXT:    retq
 ;
 ; X86-AVX512VLCDBW-LABEL: test_mm_epi64:
@@ -62,23 +59,13 @@ define <4 x i32> @test_mm_epi32(<16 x i8> %a, <16 x i8> %b) {
 ; AVX512VLCDBW-LABEL: test_mm_epi32:
 ; AVX512VLCDBW:       # BB#0: # %entry
 ; AVX512VLCDBW-NEXT:    vpcmpeqb %xmm1, %xmm0, %k0
-; AVX512VLCDBW-NEXT:    kmovd %k0, %eax
+; AVX512VLCDBW-NEXT:    vpbroadcastmw2d %k0, %xmm0
-; AVX512VLCDBW-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; AVX512VLCDBW-NEXT:    vpinsrw $0, %eax, %xmm0, %xmm0
-; AVX512VLCDBW-NEXT:    vpinsrw $2, %eax, %xmm0, %xmm0
-; AVX512VLCDBW-NEXT:    vpinsrw $4, %eax, %xmm0, %xmm0
-; AVX512VLCDBW-NEXT:    vpinsrw $6, %eax, %xmm0, %xmm0
 ; AVX512VLCDBW-NEXT:    retq
 ;
 ; X86-AVX512VLCDBW-LABEL: test_mm_epi32:
 ; X86-AVX512VLCDBW:       # BB#0: # %entry
 ; X86-AVX512VLCDBW-NEXT:    vpcmpeqb %xmm1, %xmm0, %k0
-; X86-AVX512VLCDBW-NEXT:    kmovd %k0, %eax
+; X86-AVX512VLCDBW-NEXT:    vpbroadcastmw2d %k0, %xmm0
-; X86-AVX512VLCDBW-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; X86-AVX512VLCDBW-NEXT:    vpinsrw $0, %eax, %xmm0, %xmm0
-; X86-AVX512VLCDBW-NEXT:    vpinsrw $2, %eax, %xmm0, %xmm0
-; X86-AVX512VLCDBW-NEXT:    vpinsrw $4, %eax, %xmm0, %xmm0
-; X86-AVX512VLCDBW-NEXT:    vpinsrw $6, %eax, %xmm0, %xmm0
 ; X86-AVX512VLCDBW-NEXT:    retl
 entry:
   %0 = icmp eq <16 x i8> %a, %b
@@ -100,27 +87,13 @@ define <16 x i32> @test_mm512_epi32(<16 x i32> %a, <16 x i32> %b) {
 ; AVX512VLCDBW-LABEL: test_mm512_epi32:
 ; AVX512VLCDBW:       # BB#0: # %entry
 ; AVX512VLCDBW-NEXT:    vpcmpeqd %zmm1, %zmm0, %k0
-; AVX512VLCDBW-NEXT:    kmovd %k0, %eax
+; AVX512VLCDBW-NEXT:    vpbroadcastmw2d %k0, %zmm0
-; AVX512VLCDBW-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; AVX512VLCDBW-NEXT:    vpinsrw $0, %eax, %xmm0, %xmm0
-; AVX512VLCDBW-NEXT:    vpinsrw $2, %eax, %xmm0, %xmm0
-; AVX512VLCDBW-NEXT:    vpinsrw $4, %eax, %xmm0, %xmm0
-; AVX512VLCDBW-NEXT:    vpinsrw $6, %eax, %xmm0, %xmm0
-; AVX512VLCDBW-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm0
-; AVX512VLCDBW-NEXT:    vinserti64x4 $1, %ymm0, %zmm0, %zmm0
 ; AVX512VLCDBW-NEXT:    retq
 ;
 ; X86-AVX512VLCDBW-LABEL: test_mm512_epi32:
 ; X86-AVX512VLCDBW:       # BB#0: # %entry
 ; X86-AVX512VLCDBW-NEXT:    vpcmpeqd %zmm1, %zmm0, %k0
-; X86-AVX512VLCDBW-NEXT:    kmovd %k0, %eax
+; X86-AVX512VLCDBW-NEXT:    vpbroadcastmw2d %k0, %zmm0
-; X86-AVX512VLCDBW-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; X86-AVX512VLCDBW-NEXT:    vpinsrw $0, %eax, %xmm0, %xmm0
-; X86-AVX512VLCDBW-NEXT:    vpinsrw $2, %eax, %xmm0, %xmm0
-; X86-AVX512VLCDBW-NEXT:    vpinsrw $4, %eax, %xmm0, %xmm0
-; X86-AVX512VLCDBW-NEXT:    vpinsrw $6, %eax, %xmm0, %xmm0
-; X86-AVX512VLCDBW-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm0
-; X86-AVX512VLCDBW-NEXT:    vinserti64x4 $1, %ymm0, %zmm0, %zmm0
 ; X86-AVX512VLCDBW-NEXT:    retl
 entry:
   %0 = icmp eq <16 x i32> %a, %b
@@ -145,12 +118,7 @@ define <8 x i64> @test_mm512_epi64(<8 x i32> %a, <8 x i32> %b) {
 ; AVX512VLCDBW-LABEL: test_mm512_epi64:
 ; AVX512VLCDBW:       # BB#0: # %entry
 ; AVX512VLCDBW-NEXT:    vpcmpeqd %ymm1, %ymm0, %k0
-; AVX512VLCDBW-NEXT:    kmovd %k0, %eax
+; AVX512VLCDBW-NEXT:    vpbroadcastmb2q %k0, %zmm0
-; AVX512VLCDBW-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; AVX512VLCDBW-NEXT:    vpinsrb $0, %eax, %xmm0, %xmm0
-; AVX512VLCDBW-NEXT:    vpinsrb $8, %eax, %xmm0, %xmm0
-; AVX512VLCDBW-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm0
-; AVX512VLCDBW-NEXT:    vinserti64x4 $1, %ymm0, %zmm0, %zmm0
 ; AVX512VLCDBW-NEXT:    retq
 ;
 ; X86-AVX512VLCDBW-LABEL: test_mm512_epi64:
@@ -188,11 +156,7 @@ define <4 x i64> @test_mm256_epi64(<8 x i32> %a, <8 x i32> %b) {
 ; AVX512VLCDBW-LABEL: test_mm256_epi64:
 ; AVX512VLCDBW:       # BB#0: # %entry
 ; AVX512VLCDBW-NEXT:    vpcmpeqd %ymm1, %ymm0, %k0
-; AVX512VLCDBW-NEXT:    kmovd %k0, %eax
+; AVX512VLCDBW-NEXT:    vpbroadcastmb2q %k0, %ymm0
-; AVX512VLCDBW-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; AVX512VLCDBW-NEXT:    vpinsrb $0, %eax, %xmm0, %xmm0
-; AVX512VLCDBW-NEXT:    vpinsrb $8, %eax, %xmm0, %xmm0
-; AVX512VLCDBW-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm0
 ; AVX512VLCDBW-NEXT:    retq
 ;
 ; X86-AVX512VLCDBW-LABEL: test_mm256_epi64:
@@ -232,25 +196,13 @@ define <8 x i32> @test_mm256_epi32(<16 x i16> %a, <16 x i16> %b) {
 ; AVX512VLCDBW-LABEL: test_mm256_epi32:
 ; AVX512VLCDBW:       # BB#0: # %entry
 ; AVX512VLCDBW-NEXT:    vpcmpeqw %ymm1, %ymm0, %k0
-; AVX512VLCDBW-NEXT:    kmovd %k0, %eax
+; AVX512VLCDBW-NEXT:    vpbroadcastmw2d %k0, %ymm0
-; AVX512VLCDBW-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; AVX512VLCDBW-NEXT:    vpinsrw $0, %eax, %xmm0, %xmm0
-; AVX512VLCDBW-NEXT:    vpinsrw $2, %eax, %xmm0, %xmm0
-; AVX512VLCDBW-NEXT:    vpinsrw $4, %eax, %xmm0, %xmm0
-; AVX512VLCDBW-NEXT:    vpinsrw $6, %eax, %xmm0, %xmm0
-; AVX512VLCDBW-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm0
 ; AVX512VLCDBW-NEXT:    retq
 ;
 ; X86-AVX512VLCDBW-LABEL: test_mm256_epi32:
 ; X86-AVX512VLCDBW:       # BB#0: # %entry
 ; X86-AVX512VLCDBW-NEXT:    vpcmpeqw %ymm1, %ymm0, %k0
-; X86-AVX512VLCDBW-NEXT:    kmovd %k0, %eax
+; X86-AVX512VLCDBW-NEXT:    vpbroadcastmw2d %k0, %ymm0
-; X86-AVX512VLCDBW-NEXT:    vpxor %xmm0, %xmm0, %xmm0
-; X86-AVX512VLCDBW-NEXT:    vpinsrw $0, %eax, %xmm0, %xmm0
-; X86-AVX512VLCDBW-NEXT:    vpinsrw $2, %eax, %xmm0, %xmm0
-; X86-AVX512VLCDBW-NEXT:    vpinsrw $4, %eax, %xmm0, %xmm0
-; X86-AVX512VLCDBW-NEXT:    vpinsrw $6, %eax, %xmm0, %xmm0
-; X86-AVX512VLCDBW-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm0
 ; X86-AVX512VLCDBW-NEXT:    retl
 entry:
   %0 = icmp eq <16 x i16> %a, %b