diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index d528e4921978..956f57d97f0d 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -37137,21 +37137,27 @@ static SDValue combineBitcast(SDNode *N, SelectionDAG &DAG,
     return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT,
                        DAG.getBitcast(MVT::i16, N0.getOperand(0)));
 
-  // Combine (bitcast (vbroadcast_load)) -> (vbroadcast_load). The memory VT
-  // determines the number of bits loaded.
+  // Canonicalize (bitcast (vbroadcast_load)) so that the output of the bitcast
+  // and the vbroadcast_load are both integer or both fp. In some cases this
+  // will remove the bitcast entirely.
   if (N0.getOpcode() == X86ISD::VBROADCAST_LOAD && N0.hasOneUse() &&
-      VT.getScalarSizeInBits() == SrcVT.getScalarSizeInBits()) {
+       VT.isFloatingPoint() != SrcVT.isFloatingPoint() && VT.isVector()) {
     auto *BCast = cast<MemIntrinsicSDNode>(N0);
-    assert(VT.getScalarSizeInBits() == BCast->getMemoryVT().getSizeInBits() &&
-           "Unexpected load size!");
-    SDVTList Tys = DAG.getVTList(VT, MVT::Other);
+    unsigned SrcVTSize = SrcVT.getScalarSizeInBits();
+    unsigned MemSize = BCast->getMemoryVT().getScalarSizeInBits();
+    MVT MemVT = VT.isFloatingPoint() ? MVT::getFloatingPointVT(MemSize)
+                                     : MVT::getIntegerVT(MemSize);
+    MVT LoadVT = VT.isFloatingPoint() ? MVT::getFloatingPointVT(SrcVTSize)
+                                      : MVT::getIntegerVT(SrcVTSize);
+    LoadVT = MVT::getVectorVT(LoadVT, SrcVT.getVectorNumElements());
+
+    SDVTList Tys = DAG.getVTList(LoadVT, MVT::Other);
     SDValue Ops[] = { BCast->getChain(), BCast->getBasePtr() };
     SDValue ResNode =
         DAG.getMemIntrinsicNode(X86ISD::VBROADCAST_LOAD, SDLoc(N), Tys, Ops,
-                                VT.getVectorElementType(),
-                                BCast->getMemOperand());
+                                MemVT, BCast->getMemOperand());
     DAG.ReplaceAllUsesOfValueWith(SDValue(BCast, 1), ResNode.getValue(1));
-    return ResNode;
+    return DAG.getBitcast(VT, ResNode);
   }
 
   // Since MMX types are special and don't usually play with other vector types,
diff --git a/llvm/test/CodeGen/X86/avx512dqvl-intrinsics-upgrade.ll b/llvm/test/CodeGen/X86/avx512dqvl-intrinsics-upgrade.ll
index b645098582ff..879a1643325d 100644
--- a/llvm/test/CodeGen/X86/avx512dqvl-intrinsics-upgrade.ll
+++ b/llvm/test/CodeGen/X86/avx512dqvl-intrinsics-upgrade.ll
@@ -2697,9 +2697,9 @@ define <8 x i32>@test_int_x86_avx512_mask_broadcasti32x2_256(<4 x i32> %x0, <8 x
 ; X86:       # %bb.0:
 ; X86-NEXT:    # kill: def $xmm0 killed $xmm0 def $ymm0
 ; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax # encoding: [0x8b,0x44,0x24,0x08]
-; X86-NEXT:    vpbroadcastq (%eax), %ymm2 # EVEX TO VEX Compression encoding: [0xc4,0xe2,0x7d,0x59,0x10]
 ; X86-NEXT:    kmovb {{[0-9]+}}(%esp), %k1 # encoding: [0xc5,0xf9,0x90,0x4c,0x24,0x04]
-; X86-NEXT:    vmovdqa32 %ymm2, %ymm1 {%k1} # encoding: [0x62,0xf1,0x7d,0x29,0x6f,0xca]
+; X86-NEXT:    vbroadcasti32x2 (%eax), %ymm1 {%k1} # encoding: [0x62,0xf2,0x7d,0x29,0x59,0x08]
+; X86-NEXT:    # ymm1 {%k1} = mem[0,1,0,1,0,1,0,1]
 ; X86-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm2 # EVEX TO VEX Compression encoding: [0xc4,0xe3,0x7d,0x38,0xd0,0x01]
 ; X86-NEXT:    vinserti32x4 $1, %xmm0, %ymm0, %ymm0 {%k1} {z} # encoding: [0x62,0xf3,0x7d,0xa9,0x38,0xc0,0x01]
 ; X86-NEXT:    vpaddd %ymm2, %ymm0, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xfd,0xfe,0xc2]