[X86][AVX512] lowering broadcastm intrinsic - llvm part

[jina.nahias]

Summary

This is the llvm part, the clang part: https://reviews.llvm.org/D38683

[RKSimon]

Please add fast-isel tests that match the builtin tests from D38683

[jina.nahias]

adding fast-isel tests

[RKSimon]

test/CodeGen/X86/avx512cd-intrinsics-fast-isel.ll

5

Strip the attributes

test/CodeGen/X86/avx512vlcd-intrinsics-fast-isel.ll

5

Strip the attributes

[jina.nahias]

ping

[zvi]

LGTM

[RKSimon]

test/CodeGen/X86/avx512vlcd-intrinsics-fast-isel.ll

9

Why do the avx512vlcd fast-isel cases use the vpbroadcastmb2 instructions but the avx512cd cases go through the GPRs?

[jina.nahias]

making broadcastm working without VL in 512 bit intrinsics @RKSimon @igorb @zvi

[RKSimon]

LGTM with one minor

lib/Target/X86/X86ISelLowering.cpp

6782

clang-format this

lib/IR/AutoUpgrade.cpp

static bool ShouldUpgradeX86Intrinsic(Function *F, StringRef Name) {
   // like to use this information to remove upgrade code for some older
   // intrinsics. It is currently undecided how we will determine that future
   // point.
   if (Name=="ssse3.pabs.b.128" || // Added in 6.0
       Name=="ssse3.pabs.w.128" || // Added in 6.0
       Name=="ssse3.pabs.d.128" || // Added in 6.0
       Name.startswith("avx2.pabs.") || // Added in 6.0
       Name.startswith("avx512.mask.pabs.") || // Added in 6.0
+      Name.startswith("avx512.broadcastm") || // Added in 6.0
       Name.startswith("avx512.mask.pbroadcast") || // Added in 6.0
       Name.startswith("sse2.pcmpeq.") || // Added in 3.1
       Name.startswith("sse2.pcmpgt.") || // Added in 3.1
       Name.startswith("avx2.pcmpeq.") || // Added in 3.1
       Name.startswith("avx2.pcmpgt.") || // Added in 3.1
       Name.startswith("avx512.mask.pcmpeq.") || // Added in 3.9
       Name.startswith("avx512.mask.pcmpgt.") || // Added in 3.9
       Name.startswith("avx.vperm2f128.") || // Added in 6.0
if (!NewFn) {
     // Upgrade packed integer vector compare intrinsics to compare instructions.
     if (IsX86 && (Name.startswith("sse2.pcmp") ||
                   Name.startswith("avx2.pcmp"))) {
       // "sse2.pcpmpeq." "sse2.pcmpgt." "avx2.pcmpeq." or "avx2.pcmpgt."
       bool CmpEq = Name[9] == 'e';
       Rep = Builder.CreateICmp(CmpEq ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_SGT,
                                CI->getArgOperand(0), CI->getArgOperand(1));
       Rep = Builder.CreateSExt(Rep, CI->getType(), "");
-    } else if (IsX86 && (Name.startswith("avx512.mask.pbroadcast"))){
+    } else if (IsX86 && (Name.startswith("avx512.broadcastm"))) {
+      Type *ExtTy = Type::getInt32Ty(C);
+      if (CI->getOperand(0)->getType()->isIntegerTy(8))
+        ExtTy = Type::getInt64Ty(C);
+      unsigned NumElts = CI->getType()->getPrimitiveSizeInBits() /
+                         ExtTy->getPrimitiveSizeInBits();
+      Rep = Builder.CreateZExt(CI->getArgOperand(0), ExtTy);
+      Rep = Builder.CreateVectorSplat(NumElts, Rep);
+    } else if (IsX86 && (Name.startswith("avx512.mask.pbroadcast"))) {
       unsigned NumElts =
           CI->getArgOperand(1)->getType()->getVectorNumElements();
       Rep = Builder.CreateVectorSplat(NumElts, CI->getArgOperand(0));
       Rep = EmitX86Select(Builder, CI->getArgOperand(2), Rep,
                           CI->getArgOperand(1));
     } else if (IsX86 && (Name == "sse.add.ss" || Name == "sse2.add.sd")) {
       Type *I32Ty = Type::getInt32Ty(C);
       Value *Elt0 = Builder.CreateExtractElement(CI->getArgOperand(0),

lib/Target/X86/X86ISelLowering.cpp

if (!Subtarget.hasAVX())
     return SDValue();
 
   MVT VT = BVOp->getSimpleValueType(0);
   SDLoc dl(BVOp);
 
   assert((VT.is128BitVector() || VT.is256BitVector() || VT.is512BitVector()) &&
          "Unsupported vector type for broadcast.");
 
+  BitVector UndefElements;
+  SDValue Ld = BVOp->getSplatValue(&UndefElements);
+
   // 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;
+    MVT EltType = VT.getScalarType();
+    unsigned NumElts = VT.getVectorNumElements();
+    SDValue BOperand;
     SDValue ZeroExtended = isSplatZeroExtended(BVOp, NumElts, EltType);
-    if (ZeroExtended && ZeroExtended.getOpcode() == ISD::BITCAST) {
-      SDValue BOperand = ZeroExtended.getOperand(0);
+    if ((ZeroExtended && ZeroExtended.getOpcode() == ISD::BITCAST) ||
+        (Ld && Ld.getOpcode() == ISD::ZERO_EXTEND &&
+         Ld.getOperand(0).getOpcode() == ISD::BITCAST)) {
+      if (ZeroExtended)
+        BOperand = ZeroExtended.getOperand(0);
+      else BOperand = Ld.getOperand(0).getOperand(0);
       if (BOperand.getValueType().isVector() &&
           BOperand.getSimpleValueType().getVectorElementType() == MVT::i1) {
-        if ((EltType == MVT::i64 &&
-             VT.getVectorElementType() == MVT::i8) || // for broadcastmb2q
-            (EltType == MVT::i32 &&
-             VT.getVectorElementType() == MVT::i16)) { // for broadcastmw2d
+        if ((EltType == MVT::i64 && (VT.getVectorElementType() == MVT::i8 ||
+                                     NumElts == 8)) || // for broadcastmb2q
+            (EltType == MVT::i32 && (VT.getVectorElementType() == MVT::i16 ||
+                                     NumElts == 16))) { // 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);
-
   // We need a splat of a single value to use broadcast, and it doesn't
   // make any sense if the value is only in one element of the vector.
   if (!Ld || (VT.getVectorNumElements() - UndefElements.count()) <= 1) {
     APInt SplatValue, Undef;
     unsigned SplatBitSize;
     bool HasUndef;
     // Check if this is a repeated constant pattern suitable for broadcasting.
     if (BVOp->isConstantSplat(SplatValue, Undef, SplatBitSize, HasUndef) &&

lib/Target/X86/X86IntrinsicsInfo.h

static const IntrinsicData  IntrinsicsWithoutChain[] = {
   X86_INTRINSIC_DATA(avx2_psrlv_d, INTR_TYPE_2OP, ISD::SRL, 0),
   X86_INTRINSIC_DATA(avx2_psrlv_d_256, INTR_TYPE_2OP, ISD::SRL, 0),
   X86_INTRINSIC_DATA(avx2_psrlv_q, INTR_TYPE_2OP, ISD::SRL, 0),
   X86_INTRINSIC_DATA(avx2_psrlv_q_256, INTR_TYPE_2OP, ISD::SRL, 0),
   X86_INTRINSIC_DATA(avx2_psubs_b, INTR_TYPE_2OP, X86ISD::SUBS, 0),
   X86_INTRINSIC_DATA(avx2_psubs_w, INTR_TYPE_2OP, X86ISD::SUBS, 0),
   X86_INTRINSIC_DATA(avx2_psubus_b, INTR_TYPE_2OP, X86ISD::SUBUS, 0),
   X86_INTRINSIC_DATA(avx2_psubus_w, INTR_TYPE_2OP, X86ISD::SUBUS, 0),
-  X86_INTRINSIC_DATA(avx512_broadcastmb_128, BROADCASTM, X86ISD::VBROADCASTM, 0),
-  X86_INTRINSIC_DATA(avx512_broadcastmb_256, BROADCASTM, X86ISD::VBROADCASTM, 0),
-  X86_INTRINSIC_DATA(avx512_broadcastmb_512, BROADCASTM, X86ISD::VBROADCASTM, 0),
-  X86_INTRINSIC_DATA(avx512_broadcastmw_128, BROADCASTM, X86ISD::VBROADCASTM, 0),
-  X86_INTRINSIC_DATA(avx512_broadcastmw_256, BROADCASTM, X86ISD::VBROADCASTM, 0),
-  X86_INTRINSIC_DATA(avx512_broadcastmw_512, BROADCASTM, X86ISD::VBROADCASTM, 0),
   X86_INTRINSIC_DATA(avx512_cvtb2mask_128, CONVERT_TO_MASK, X86ISD::CVT2MASK, 0),
   X86_INTRINSIC_DATA(avx512_cvtb2mask_256, CONVERT_TO_MASK, X86ISD::CVT2MASK, 0),
   X86_INTRINSIC_DATA(avx512_cvtb2mask_512, CONVERT_TO_MASK, X86ISD::CVT2MASK, 0),
   X86_INTRINSIC_DATA(avx512_cvtd2mask_128, CONVERT_TO_MASK, X86ISD::CVT2MASK, 0),
   X86_INTRINSIC_DATA(avx512_cvtd2mask_256, CONVERT_TO_MASK, X86ISD::CVT2MASK, 0),
   X86_INTRINSIC_DATA(avx512_cvtd2mask_512, CONVERT_TO_MASK, X86ISD::CVT2MASK, 0),
   X86_INTRINSIC_DATA(avx512_cvtq2mask_128, CONVERT_TO_MASK, X86ISD::CVT2MASK, 0),
   X86_INTRINSIC_DATA(avx512_cvtq2mask_256, CONVERT_TO_MASK, X86ISD::CVT2MASK, 0),

test/CodeGen/X86/avx512cd-intrinsics-fast-isel.ll

+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc < %s -fast-isel -mtriple=x86_64-unknown-unknown -mattr=+avx512cd | FileCheck %s
+
+define <8 x i64> @test_mm512_broadcastmb_epi64(<8 x i64> %a, <8 x i64> %b) {
+; CHECK-LABEL: test_mm512_broadcastmb_epi64:
+; CHECK:       # BB#0: # %entry
+; CHECK-NEXT:    vpcmpeqq %zmm1, %zmm0, %k0
+; CHECK-NEXT:    vpbroadcastmb2q %k0, %zmm0
+; CHECK-NEXT:    retq
+entry:
+  %0 = icmp eq <8 x i64> %a, %b
+  %1 = bitcast <8 x i1> %0 to i8
+  %conv.i = zext i8 %1 to i64
+  %vecinit.i.i = insertelement <8 x i64> undef, i64 %conv.i, i32 0
+  %vecinit7.i.i = shufflevector <8 x i64> %vecinit.i.i, <8 x i64> undef, <8 x i32> zeroinitializer
+  ret <8 x i64> %vecinit7.i.i
+}
+
+define <8 x i64> @test_mm512_broadcastmw_epi32(<8 x i64> %a, <8 x i64> %b) {
+; CHECK-LABEL: test_mm512_broadcastmw_epi32:
+; CHECK:       # BB#0: # %entry
+; CHECK-NEXT:    vpcmpeqd %zmm1, %zmm0, %k0
+; CHECK-NEXT:    vpbroadcastmw2d %k0, %zmm0
+; CHECK-NEXT:    retq
+entry:
+  %0 = bitcast <8 x i64> %a to <16 x i32>
+  %1 = bitcast <8 x i64> %b to <16 x i32>
+  %2 = icmp eq <16 x i32> %0, %1
+  %3 = bitcast <16 x i1> %2 to i16
+  %conv.i = zext i16 %3 to i32
+  %vecinit.i.i = insertelement <16 x i32> undef, i32 %conv.i, i32 0
+  %vecinit15.i.i = shufflevector <16 x i32> %vecinit.i.i, <16 x i32> undef, <16 x i32> zeroinitializer
+  %4 = bitcast <16 x i32> %vecinit15.i.i to <8 x i64>
+  ret <8 x i64> %4
+}
+
+

test/CodeGen/X86/avx512cd-intrinsics-upgrade.ll

 ; CHECK:       ## BB#0:
 ; CHECK-NEXT:    kmovw %edi, %k1
 ; CHECK-NEXT:    vplzcntq %zmm0, %zmm1 {%k1}
 ; CHECK-NEXT:    vmovdqa64 %zmm1, %zmm0
 ; CHECK-NEXT:    retq
   %res = call <8 x i64> @llvm.x86.avx512.mask.lzcnt.q.512(<8 x i64> %a, <8 x i64> %b, i8 %mask)
   ret <8 x i64> %res
 }
+
+define <16 x i32> @test_x86_vbroadcastmw_512(i16 %a0) {
+; CHECK-LABEL: test_x86_vbroadcastmw_512:
+; CHECK:       ## BB#0:
+; CHECK-NEXT:    movzwl %di, %eax
+; CHECK-NEXT:    vpbroadcastd %eax, %zmm0
+; CHECK-NEXT:    retq
+  %res = call <16 x i32> @llvm.x86.avx512.broadcastmw.512(i16 %a0)
+  ret <16 x i32> %res
+}
+declare <16 x i32> @llvm.x86.avx512.broadcastmw.512(i16)
+
+define <8 x i64> @test_x86_broadcastmb_512(i8 %a0) {
+; CHECK-LABEL: test_x86_broadcastmb_512:
+; CHECK:       ## BB#0:
+; CHECK-NEXT:    movzbl %dil, %eax
+; CHECK-NEXT:    vpbroadcastq %rax, %zmm0
+; CHECK-NEXT:    retq
+  %res = call <8 x i64> @llvm.x86.avx512.broadcastmb.512(i8 %a0)
+  ret <8 x i64> %res
+}
+declare <8 x i64> @llvm.x86.avx512.broadcastmb.512(i8)
+

test/CodeGen/X86/avx512cd-intrinsics.ll

 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
 ; RUN: llc < %s -mtriple=x86_64-apple-darwin -mcpu=knl -mattr=+avx512cd | FileCheck %s
 
-define <16 x i32> @test_x86_vbroadcastmw_512(i16 %a0) {
-; CHECK-LABEL: test_x86_vbroadcastmw_512:
-; CHECK:       ## BB#0:
-; CHECK-NEXT:    kmovw %edi, %k0
-; CHECK-NEXT:    vpbroadcastmw2d %k0, %zmm0
-; CHECK-NEXT:    retq
-  %res = call <16 x i32> @llvm.x86.avx512.broadcastmw.512(i16 %a0)
-  ret <16 x i32> %res
-}
-declare <16 x i32> @llvm.x86.avx512.broadcastmw.512(i16)
-
-define <8 x i64> @test_x86_broadcastmb_512(i8 %a0) {
-; CHECK-LABEL: test_x86_broadcastmb_512:
-; CHECK:       ## BB#0:
-; CHECK-NEXT:    kmovw %edi, %k0
-; CHECK-NEXT:    vpbroadcastmb2q %k0, %zmm0
-; CHECK-NEXT:    retq
-  %res = call <8 x i64> @llvm.x86.avx512.broadcastmb.512(i8 %a0)
-  ret <8 x i64> %res
-}
-declare <8 x i64> @llvm.x86.avx512.broadcastmb.512(i8)
-
 declare <16 x i32> @llvm.x86.avx512.mask.conflict.d.512(<16 x i32>, <16 x i32>, i16) nounwind readonly
 
 define <8 x i64> @test_conflict_q(<8 x i64> %a) {
 ; CHECK-LABEL: test_conflict_q:
 ; CHECK:       ## BB#0:
 ; CHECK-NEXT:    vpconflictq %zmm0, %zmm0
 ; CHECK-NEXT:    retq
   %res = call <8 x i64> @llvm.x86.avx512.mask.conflict.q.512(<8 x i64> %a, <8 x i64> zeroinitializer, i8 -1)

test/CodeGen/X86/avx512cdvl-intrinsics-upgrade.ll

 ; CHECK-NEXT:    vpaddq %ymm2, %ymm1, %ymm0
 ; CHECK-NEXT:    retq
   %res = call <4 x i64> @llvm.x86.avx512.mask.lzcnt.q.256(<4 x i64> %x0, <4 x i64> %x1, i8 %x2)
   %res1 = call <4 x i64> @llvm.x86.avx512.mask.lzcnt.q.256(<4 x i64> %x0, <4 x i64> %x1, i8 -1)
   %res2 = add <4 x i64> %res, %res1
   ret <4 x i64> %res2
 }
 
+define <8 x i32> @test_x86_vbroadcastmw_256(i16 %a0) {
+; CHECK-LABEL: test_x86_vbroadcastmw_256:
+; CHECK:       ## BB#0:
+; CHECK-NEXT:    movzwl %di, %eax
+; CHECK-NEXT:    vpbroadcastd %eax, %ymm0
+; CHECK-NEXT:    retq
+  %res = call <8 x i32> @llvm.x86.avx512.broadcastmw.256(i16 %a0) ;
+  ret <8 x i32> %res
+}
+declare <8 x i32> @llvm.x86.avx512.broadcastmw.256(i16)
+
+define <4 x i32> @test_x86_vbroadcastmw_128(i16 %a0) {
+; CHECK-LABEL: test_x86_vbroadcastmw_128:
+; CHECK:       ## BB#0:
+; CHECK-NEXT:    movzwl %di, %eax
+; CHECK-NEXT:    vpbroadcastd %eax, %xmm0
+; CHECK-NEXT:    retq
+  %res = call <4 x i32> @llvm.x86.avx512.broadcastmw.128(i16 %a0) ;
+  ret <4 x i32> %res
+}
+declare <4 x i32> @llvm.x86.avx512.broadcastmw.128(i16)
+
+define <4 x i64> @test_x86_broadcastmb_256(i8 %a0) {
+; CHECK-LABEL: test_x86_broadcastmb_256:
+; CHECK:       ## BB#0:
+; CHECK-NEXT:    movzbl %dil, %eax
+; CHECK-NEXT:    vpbroadcastq %rax, %ymm0
+; CHECK-NEXT:    retq
+  %res = call <4 x i64> @llvm.x86.avx512.broadcastmb.256(i8 %a0) ;
+  ret <4 x i64> %res
+}
+declare <4 x i64> @llvm.x86.avx512.broadcastmb.256(i8)
+
+define <2 x i64> @test_x86_broadcastmb_128(i8 %a0) {
+; CHECK-LABEL: test_x86_broadcastmb_128:
+; CHECK:       ## BB#0:
+; CHECK-NEXT:    movzbl %dil, %eax
+; CHECK-NEXT:    vpbroadcastq %rax, %xmm0
+; CHECK-NEXT:    retq
+  %res = call <2 x i64> @llvm.x86.avx512.broadcastmb.128(i8 %a0) ;
+  ret <2 x i64> %res
+}
+declare <2 x i64> @llvm.x86.avx512.broadcastmb.128(i8)
+

test/CodeGen/X86/avx512cdvl-intrinsics.ll

 ; CHECK-NEXT:    vpaddq %ymm0, %ymm1, %ymm0
 ; CHECK-NEXT:    retq
   %res = call <4 x i64> @llvm.x86.avx512.mask.conflict.q.256(<4 x i64> %x0, <4 x i64> %x1, i8 %x2)
   %res1 = call <4 x i64> @llvm.x86.avx512.mask.conflict.q.256(<4 x i64> %x0, <4 x i64> %x1, i8 -1)
   %res2 = add <4 x i64> %res, %res1
   ret <4 x i64> %res2
 }
 
-define <8 x i32> @test_x86_vbroadcastmw_256(i16 %a0) {
-; CHECK-LABEL: test_x86_vbroadcastmw_256:
-; CHECK:       ## BB#0:
-; CHECK-NEXT:    kmovw %edi, %k0
-; CHECK-NEXT:    vpbroadcastmw2d %k0, %ymm0
-; CHECK-NEXT:    retq
-  %res = call <8 x i32> @llvm.x86.avx512.broadcastmw.256(i16 %a0) ;
-  ret <8 x i32> %res
-}
-declare <8 x i32> @llvm.x86.avx512.broadcastmw.256(i16)
-
-define <4 x i32> @test_x86_vbroadcastmw_128(i16 %a0) {
-; CHECK-LABEL: test_x86_vbroadcastmw_128:
-; CHECK:       ## BB#0:
-; CHECK-NEXT:    kmovw %edi, %k0
-; CHECK-NEXT:    vpbroadcastmw2d %k0, %xmm0
-; CHECK-NEXT:    retq
-  %res = call <4 x i32> @llvm.x86.avx512.broadcastmw.128(i16 %a0) ;
-  ret <4 x i32> %res
-}
-declare <4 x i32> @llvm.x86.avx512.broadcastmw.128(i16)
-
-define <4 x i64> @test_x86_broadcastmb_256(i8 %a0) {
-; CHECK-LABEL: test_x86_broadcastmb_256:
-; CHECK:       ## BB#0:
-; CHECK-NEXT:    kmovw %edi, %k0
-; CHECK-NEXT:    vpbroadcastmb2q %k0, %ymm0
-; CHECK-NEXT:    retq
-  %res = call <4 x i64> @llvm.x86.avx512.broadcastmb.256(i8 %a0) ;
-  ret <4 x i64> %res
-}
-declare <4 x i64> @llvm.x86.avx512.broadcastmb.256(i8)
-
-define <2 x i64> @test_x86_broadcastmb_128(i8 %a0) {
-; CHECK-LABEL: test_x86_broadcastmb_128:
-; CHECK:       ## BB#0:
-; CHECK-NEXT:    kmovw %edi, %k0
-; CHECK-NEXT:    vpbroadcastmb2q %k0, %xmm0
-; CHECK-NEXT:    retq
-  %res = call <2 x i64> @llvm.x86.avx512.broadcastmb.128(i8 %a0) ;
-  ret <2 x i64> %res
-}
-declare <2 x i64> @llvm.x86.avx512.broadcastmb.128(i8)

test/CodeGen/X86/avx512vlcd-intrinsics-fast-isel.ll

+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc < %s -fast-isel -mtriple=x86_64-unknown-unknown -mattr=+avx512cd,+avx512vl | FileCheck %s
+
+define <2 x i64> @test_mm_broadcastmb_epi64(<2 x i64> %a, <2 x i64> %b) {
+; CHECK-LABEL: test_mm_broadcastmb_epi64:
+; CHECK:       # BB#0: # %entry
+; CHECK-NEXT:    vpcmpeqd %xmm1, %xmm0, %k0
+; CHECK-NEXT:    vpbroadcastmb2q %k0, %xmm0
+; CHECK-NEXT:    retq
+entry:
+  %0 = bitcast <2 x i64> %a to <4 x i32>
+  %1 = bitcast <2 x i64> %b to <4 x i32>
+  %2 = icmp eq <4 x i32> %0, %1
+  %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
+  %4 = bitcast <8 x i1> %3 to i8
+  %conv.i = zext i8 %4 to i64
+  %vecinit.i.i = insertelement <2 x i64> undef, i64 %conv.i, i32 0
+  %vecinit1.i.i = shufflevector <2 x i64> %vecinit.i.i, <2 x i64> undef, <2 x i32> zeroinitializer
+  ret <2 x i64> %vecinit1.i.i
+}
+
+define <4 x i64> @test_mm256_broadcastmb_epi64(<4 x i64> %a, <4 x i64> %b) {
+; CHECK-LABEL: test_mm256_broadcastmb_epi64:
+; CHECK:       # BB#0: # %entry
+; CHECK-NEXT:    vpcmpeqq %ymm1, %ymm0, %k0
+; CHECK-NEXT:    vpbroadcastmb2q %k0, %ymm0
+; CHECK-NEXT:    retq
+entry:
+  %0 = icmp eq <4 x i64> %a, %b
+  %1 = shufflevector <4 x i1> %0, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
+  %2 = bitcast <8 x i1> %1 to i8
+  %conv.i = zext i8 %2 to i64
+  %vecinit.i.i = insertelement <4 x i64> undef, i64 %conv.i, i32 0
+  %vecinit3.i.i = shufflevector <4 x i64> %vecinit.i.i, <4 x i64> undef, <4 x i32> zeroinitializer
+  ret <4 x i64> %vecinit3.i.i
+}
+
+define <2 x i64> @test_mm_broadcastmw_epi32(<8 x i64> %a, <8 x i64> %b) {
+; CHECK-LABEL: test_mm_broadcastmw_epi32:
+; CHECK:       # BB#0: # %entry
+; CHECK-NEXT:    vpcmpeqd %zmm1, %zmm0, %k0
+; CHECK-NEXT:    vpbroadcastmw2d %k0, %xmm0
+; CHECK-NEXT:    vzeroupper
+; CHECK-NEXT:    retq
+entry:
+  %0 = bitcast <8 x i64> %a to <16 x i32>
+  %1 = bitcast <8 x i64> %b to <16 x i32>
+  %2 = icmp eq <16 x i32> %0, %1
+  %3 = bitcast <16 x i1> %2 to i16
+  %conv.i = zext i16 %3 to i32
+  %vecinit.i.i = insertelement <4 x i32> undef, i32 %conv.i, i32 0
+  %vecinit3.i.i = shufflevector <4 x i32> %vecinit.i.i, <4 x i32> undef, <4 x i32> zeroinitializer
+  %4 = bitcast <4 x i32> %vecinit3.i.i to <2 x i64>
+  ret <2 x i64> %4
+}
+
+define <4 x i64> @test_mm256_broadcastmw_epi32(<8 x i64> %a, <8 x i64> %b) {
+; CHECK-LABEL: test_mm256_broadcastmw_epi32:
+; CHECK:       # BB#0: # %entry
+; CHECK-NEXT:    vpcmpeqd %zmm1, %zmm0, %k0
+; CHECK-NEXT:    vpbroadcastmw2d %k0, %ymm0
+; CHECK-NEXT:    retq
+entry:
+  %0 = bitcast <8 x i64> %a to <16 x i32>
+  %1 = bitcast <8 x i64> %b to <16 x i32>
+  %2 = icmp eq <16 x i32> %0, %1
+  %3 = bitcast <16 x i1> %2 to i16
+  %conv.i = zext i16 %3 to i32
+  %vecinit.i.i = insertelement <8 x i32> undef, i32 %conv.i, i32 0
+  %vecinit7.i.i = shufflevector <8 x i32> %vecinit.i.i, <8 x i32> undef, <8 x i32> zeroinitializer
+  %4 = bitcast <8 x i32> %vecinit7.i.i to <4 x i64>
+  ret <4 x i64> %4
+}
+
+

test/CodeGen/X86/broadcastm-lowering.ll

entry:
   %vecinit3.i.i = shufflevector <4 x i32> %vecinit.i.i, <4 x i32> undef, <4 x i32> zeroinitializer
   ret <4 x i32> %vecinit3.i.i
 }
 
 define <16 x i32> @test_mm512_epi32(<16 x i32> %a, <16 x i32> %b) {
 ; AVX512CD-LABEL: test_mm512_epi32:
 ; AVX512CD:       # BB#0: # %entry
 ; AVX512CD-NEXT:    vpcmpeqd %zmm1, %zmm0, %k0
-; AVX512CD-NEXT:    kmovw %k0, %eax
-; AVX512CD-NEXT:    vpbroadcastd %eax, %zmm0
+; AVX512CD-NEXT:    vpbroadcastmw2d %k0, %zmm0
 ; AVX512CD-NEXT:    retq
 ;
 ; AVX512VLCDBW-LABEL: test_mm512_epi32:
 ; AVX512VLCDBW:       # BB#0: # %entry
 ; AVX512VLCDBW-NEXT:    vpcmpeqd %zmm1, %zmm0, %k0
 ; AVX512VLCDBW-NEXT:    vpbroadcastmw2d %k0, %zmm0
 ; AVX512VLCDBW-NEXT:    retq
 ;

 }
 
 define <8 x i64> @test_mm512_epi64(<8 x i32> %a, <8 x i32> %b) {
 ; AVX512CD-LABEL: test_mm512_epi64:
 ; AVX512CD:       # BB#0: # %entry
 ; AVX512CD-NEXT:    # kill: %YMM1<def> %YMM1<kill> %ZMM1<def>
 ; AVX512CD-NEXT:    # kill: %YMM0<def> %YMM0<kill> %ZMM0<def>
 ; AVX512CD-NEXT:    vpcmpeqd %zmm1, %zmm0, %k0
-; AVX512CD-NEXT:    kmovw %k0, %eax
-; AVX512CD-NEXT:    movzbl %al, %eax
-; AVX512CD-NEXT:    vpbroadcastq %rax, %zmm0
+; AVX512CD-NEXT:    vpbroadcastmb2q %k0, %zmm0
 ; AVX512CD-NEXT:    retq
 ;
 ; AVX512VLCDBW-LABEL: test_mm512_epi64:
 ; AVX512VLCDBW:       # BB#0: # %entry
 ; AVX512VLCDBW-NEXT:    vpcmpeqd %ymm1, %ymm0, %k0
 ; AVX512VLCDBW-NEXT:    vpbroadcastmb2q %k0, %zmm0
 ; AVX512VLCDBW-NEXT:    retq
 ;