diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index dc46d0e499e..0bc6de0af79 100644
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -1547,6 +1547,7 @@ X86TargetLowering::X86TargetLowering(const X86TargetMachine &TM,
     setOperationAction(ISD::SUB,                MVT::v32i16, Legal);
     setOperationAction(ISD::SUB,                MVT::v64i8, Legal);
     setOperationAction(ISD::MUL,                MVT::v32i16, Legal);
+    setOperationAction(ISD::MUL,                MVT::v64i8, Custom);
     setOperationAction(ISD::MULHS,              MVT::v32i16, Legal);
     setOperationAction(ISD::MULHU,              MVT::v32i16, Legal);
     setOperationAction(ISD::CONCAT_VECTORS,     MVT::v32i1, Custom);
@@ -18967,6 +18968,35 @@ static SDValue Lower256IntArith(SDValue Op, SelectionDAG &DAG) {
                      DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, RHS2));
 }
 
+/// Break a 512-bit integer operation into two new 256-bit ones and then
+/// concatenate the result back.
+static SDValue Lower512IntArith(SDValue Op, SelectionDAG &DAG) {
+  MVT VT = Op.getSimpleValueType();
+
+  assert(VT.is512BitVector() && VT.isInteger() &&
+         "Unsupported value type for operation");
+
+  unsigned NumElems = VT.getVectorNumElements();
+  SDLoc dl(Op);
+
+  // Extract the LHS vectors
+  SDValue LHS = Op.getOperand(0);
+  SDValue LHS1 = extract256BitVector(LHS, 0, DAG, dl);
+  SDValue LHS2 = extract256BitVector(LHS, NumElems / 2, DAG, dl);
+
+  // Extract the RHS vectors
+  SDValue RHS = Op.getOperand(1);
+  SDValue RHS1 = extract256BitVector(RHS, 0, DAG, dl);
+  SDValue RHS2 = extract256BitVector(RHS, NumElems / 2, DAG, dl);
+
+  MVT EltVT = VT.getVectorElementType();
+  MVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
+
+  return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
+                     DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, RHS1),
+                     DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, RHS2));
+}
+
 static SDValue LowerADD(SDValue Op, SelectionDAG &DAG) {
   if (Op.getValueType() == MVT::i1)
     return DAG.getNode(ISD::XOR, SDLoc(Op), Op.getValueType(),
@@ -19009,10 +19039,15 @@ static SDValue LowerMUL(SDValue Op, const X86Subtarget &Subtarget,
   SDValue A = Op.getOperand(0);
   SDValue B = Op.getOperand(1);
 
-  // Lower v16i8/v32i8 mul as sign-extension to v8i16/v16i16 vector
-  // pairs, multiply and truncate.
-  if (VT == MVT::v16i8 || VT == MVT::v32i8) {
+  // Lower v16i8/v32i8/v64i8 mul as sign-extension to v8i16/v16i16/v32i16
+  // vector pairs, multiply and truncate.
+  if (VT == MVT::v16i8 || VT == MVT::v32i8 || VT == MVT::v64i8) {
     if (Subtarget.hasInt256()) {
+      // For 512-bit vectors, split into 256-bit vectors to allow the
+      // sign-extension to occur.
+      if (VT == MVT::v64i8)
+        return Lower512IntArith(Op, DAG);
+
       // For 256-bit vectors, split into 128-bit vectors to allow the
       // sign-extension to occur. We don't need this on AVX512BW as we can
       // safely sign-extend to v32i16.
diff --git a/test/CodeGen/X86/pmul.ll b/test/CodeGen/X86/pmul.ll
index 7370aa31fe6..657d51c05bd 100644
--- a/test/CodeGen/X86/pmul.ll
+++ b/test/CodeGen/X86/pmul.ll
@@ -929,269 +929,16 @@ define <64 x i8> @mul_v64i8c(<64 x i8> %i) nounwind  {
 ;
 ; AVX512BW-LABEL: mul_v64i8c:
 ; AVX512BW:       # BB#0: # %entry
-; AVX512BW-NEXT:    vextracti32x4 $3, %zmm0, %xmm1
-; AVX512BW-NEXT:    vpextrb $1, %xmm1, %eax
-; AVX512BW-NEXT:    movb $117, %cl
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %edx
-; AVX512BW-NEXT:    vpextrb $0, %xmm1, %eax
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vmovd %eax, %xmm2
-; AVX512BW-NEXT:    vpinsrb $1, %edx, %xmm2, %xmm2
-; AVX512BW-NEXT:    vpextrb $2, %xmm1, %eax
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $2, %eax, %xmm2, %xmm2
-; AVX512BW-NEXT:    vpextrb $3, %xmm1, %eax
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $3, %eax, %xmm2, %xmm2
-; AVX512BW-NEXT:    vpextrb $4, %xmm1, %eax
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $4, %eax, %xmm2, %xmm2
-; AVX512BW-NEXT:    vpextrb $5, %xmm1, %eax
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $5, %eax, %xmm2, %xmm2
-; AVX512BW-NEXT:    vpextrb $6, %xmm1, %eax
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $6, %eax, %xmm2, %xmm2
-; AVX512BW-NEXT:    vpextrb $7, %xmm1, %eax
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $7, %eax, %xmm2, %xmm2
-; AVX512BW-NEXT:    vpextrb $8, %xmm1, %eax
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $8, %eax, %xmm2, %xmm2
-; AVX512BW-NEXT:    vpextrb $9, %xmm1, %eax
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $9, %eax, %xmm2, %xmm2
-; AVX512BW-NEXT:    vpextrb $10, %xmm1, %eax
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $10, %eax, %xmm2, %xmm2
-; AVX512BW-NEXT:    vpextrb $11, %xmm1, %eax
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $11, %eax, %xmm2, %xmm2
-; AVX512BW-NEXT:    vpextrb $12, %xmm1, %eax
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $12, %eax, %xmm2, %xmm2
-; AVX512BW-NEXT:    vpextrb $13, %xmm1, %eax
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $13, %eax, %xmm2, %xmm2
-; AVX512BW-NEXT:    vpextrb $14, %xmm1, %eax
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $14, %eax, %xmm2, %xmm2
-; AVX512BW-NEXT:    vpextrb $15, %xmm1, %eax
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $15, %eax, %xmm2, %xmm1
-; AVX512BW-NEXT:    vextracti32x4 $2, %zmm0, %xmm2
-; AVX512BW-NEXT:    vpextrb $1, %xmm2, %eax
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %edx
-; AVX512BW-NEXT:    vpextrb $0, %xmm2, %eax
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vmovd %eax, %xmm3
-; AVX512BW-NEXT:    vpinsrb $1, %edx, %xmm3, %xmm3
-; AVX512BW-NEXT:    vpextrb $2, %xmm2, %eax
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $2, %eax, %xmm3, %xmm3
-; AVX512BW-NEXT:    vpextrb $3, %xmm2, %eax
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $3, %eax, %xmm3, %xmm3
-; AVX512BW-NEXT:    vpextrb $4, %xmm2, %eax
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $4, %eax, %xmm3, %xmm3
-; AVX512BW-NEXT:    vpextrb $5, %xmm2, %eax
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $5, %eax, %xmm3, %xmm3
-; AVX512BW-NEXT:    vpextrb $6, %xmm2, %eax
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $6, %eax, %xmm3, %xmm3
-; AVX512BW-NEXT:    vpextrb $7, %xmm2, %eax
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $7, %eax, %xmm3, %xmm3
-; AVX512BW-NEXT:    vpextrb $8, %xmm2, %eax
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $8, %eax, %xmm3, %xmm3
-; AVX512BW-NEXT:    vpextrb $9, %xmm2, %eax
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $9, %eax, %xmm3, %xmm3
-; AVX512BW-NEXT:    vpextrb $10, %xmm2, %eax
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $10, %eax, %xmm3, %xmm3
-; AVX512BW-NEXT:    vpextrb $11, %xmm2, %eax
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $11, %eax, %xmm3, %xmm3
-; AVX512BW-NEXT:    vpextrb $12, %xmm2, %eax
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $12, %eax, %xmm3, %xmm3
-; AVX512BW-NEXT:    vpextrb $13, %xmm2, %eax
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $13, %eax, %xmm3, %xmm3
-; AVX512BW-NEXT:    vpextrb $14, %xmm2, %eax
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $14, %eax, %xmm3, %xmm3
-; AVX512BW-NEXT:    vpextrb $15, %xmm2, %eax
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $15, %eax, %xmm3, %xmm2
-; AVX512BW-NEXT:    vinserti128 $1, %xmm1, %ymm2, %ymm1
-; AVX512BW-NEXT:    vextracti32x4 $1, %zmm0, %xmm2
-; AVX512BW-NEXT:    vpextrb $1, %xmm2, %eax
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %edx
-; AVX512BW-NEXT:    vpextrb $0, %xmm2, %eax
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vmovd %eax, %xmm3
-; AVX512BW-NEXT:    vpinsrb $1, %edx, %xmm3, %xmm3
-; AVX512BW-NEXT:    vpextrb $2, %xmm2, %eax
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $2, %eax, %xmm3, %xmm3
-; AVX512BW-NEXT:    vpextrb $3, %xmm2, %eax
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $3, %eax, %xmm3, %xmm3
-; AVX512BW-NEXT:    vpextrb $4, %xmm2, %eax
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $4, %eax, %xmm3, %xmm3
-; AVX512BW-NEXT:    vpextrb $5, %xmm2, %eax
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $5, %eax, %xmm3, %xmm3
-; AVX512BW-NEXT:    vpextrb $6, %xmm2, %eax
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $6, %eax, %xmm3, %xmm3
-; AVX512BW-NEXT:    vpextrb $7, %xmm2, %eax
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $7, %eax, %xmm3, %xmm3
-; AVX512BW-NEXT:    vpextrb $8, %xmm2, %eax
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $8, %eax, %xmm3, %xmm3
-; AVX512BW-NEXT:    vpextrb $9, %xmm2, %eax
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $9, %eax, %xmm3, %xmm3
-; AVX512BW-NEXT:    vpextrb $10, %xmm2, %eax
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $10, %eax, %xmm3, %xmm3
-; AVX512BW-NEXT:    vpextrb $11, %xmm2, %eax
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $11, %eax, %xmm3, %xmm3
-; AVX512BW-NEXT:    vpextrb $12, %xmm2, %eax
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $12, %eax, %xmm3, %xmm3
-; AVX512BW-NEXT:    vpextrb $13, %xmm2, %eax
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $13, %eax, %xmm3, %xmm3
-; AVX512BW-NEXT:    vpextrb $14, %xmm2, %eax
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $14, %eax, %xmm3, %xmm3
-; AVX512BW-NEXT:    vpextrb $15, %xmm2, %eax
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $15, %eax, %xmm3, %xmm2
-; AVX512BW-NEXT:    vpextrb $1, %xmm0, %eax
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %edx
-; AVX512BW-NEXT:    vpextrb $0, %xmm0, %eax
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vmovd %eax, %xmm3
-; AVX512BW-NEXT:    vpinsrb $1, %edx, %xmm3, %xmm3
-; AVX512BW-NEXT:    vpextrb $2, %xmm0, %eax
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $2, %eax, %xmm3, %xmm3
-; AVX512BW-NEXT:    vpextrb $3, %xmm0, %eax
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $3, %eax, %xmm3, %xmm3
-; AVX512BW-NEXT:    vpextrb $4, %xmm0, %eax
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $4, %eax, %xmm3, %xmm3
-; AVX512BW-NEXT:    vpextrb $5, %xmm0, %eax
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $5, %eax, %xmm3, %xmm3
-; AVX512BW-NEXT:    vpextrb $6, %xmm0, %eax
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $6, %eax, %xmm3, %xmm3
-; AVX512BW-NEXT:    vpextrb $7, %xmm0, %eax
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $7, %eax, %xmm3, %xmm3
-; AVX512BW-NEXT:    vpextrb $8, %xmm0, %eax
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $8, %eax, %xmm3, %xmm3
-; AVX512BW-NEXT:    vpextrb $9, %xmm0, %eax
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $9, %eax, %xmm3, %xmm3
-; AVX512BW-NEXT:    vpextrb $10, %xmm0, %eax
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $10, %eax, %xmm3, %xmm3
-; AVX512BW-NEXT:    vpextrb $11, %xmm0, %eax
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $11, %eax, %xmm3, %xmm3
-; AVX512BW-NEXT:    vpextrb $12, %xmm0, %eax
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $12, %eax, %xmm3, %xmm3
-; AVX512BW-NEXT:    vpextrb $13, %xmm0, %eax
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $13, %eax, %xmm3, %xmm3
-; AVX512BW-NEXT:    vpextrb $14, %xmm0, %eax
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $14, %eax, %xmm3, %xmm3
-; AVX512BW-NEXT:    vpextrb $15, %xmm0, %eax
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $15, %eax, %xmm3, %xmm0
-; AVX512BW-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm0
-; AVX512BW-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
+; AVX512BW-NEXT:    vmovaps {{.*#+}} ymm1 = [117,117,117,117,117,117,117,117,117,117,117,117,117,117,117,117,117,117,117,117,117,117,117,117,117,117,117,117,117,117,117,117]
+; AVX512BW-NEXT:    vpmovsxbw %ymm1, %zmm1
+; AVX512BW-NEXT:    vpmovsxbw %ymm0, %zmm2
+; AVX512BW-NEXT:    vpmullw %zmm1, %zmm2, %zmm2
+; AVX512BW-NEXT:    vpmovwb %zmm2, %ymm2
+; AVX512BW-NEXT:    vextracti64x4 $1, %zmm0, %ymm0
+; AVX512BW-NEXT:    vpmovsxbw %ymm0, %zmm0
+; AVX512BW-NEXT:    vpmullw %zmm1, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpmovwb %zmm0, %ymm0
+; AVX512BW-NEXT:    vinserti64x4 $1, %ymm0, %zmm2, %zmm0
 ; AVX512BW-NEXT:    retq
 entry:
   %A = mul <64 x i8> %i, < i8 117, i8 117, i8 117, i8 117, i8 117, i8 117, i8 117, i8 117, i8 117, i8 117, i8 117, i8 117, i8 117, i8 117, i8 117, i8 117, i8 117, i8 117, i8 117, i8 117, i8 117, i8 117, i8 117, i8 117, i8 117, i8 117, i8 117, i8 117, i8 117, i8 117, i8 117, i8 117, i8 117, i8 117, i8 117, i8 117, i8 117, i8 117, i8 117, i8 117, i8 117, i8 117, i8 117, i8 117, i8 117, i8 117, i8 117, i8 117, i8 117, i8 117, i8 117, i8 117, i8 117, i8 117, i8 117, i8 117, i8 117, i8 117, i8 117, i8 117, i8 117, i8 117, i8 117, i8 117 >
@@ -1388,335 +1135,17 @@ define <64 x i8> @mul_v64i8(<64 x i8> %i, <64 x i8> %j) nounwind  {
 ;
 ; AVX512BW-LABEL: mul_v64i8:
 ; AVX512BW:       # BB#0: # %entry
-; AVX512BW-NEXT:    vextracti32x4 $3, %zmm0, %xmm2
-; AVX512BW-NEXT:    vpextrb $1, %xmm2, %eax
-; AVX512BW-NEXT:    vextracti32x4 $3, %zmm1, %xmm3
-; AVX512BW-NEXT:    vpextrb $1, %xmm3, %ecx
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %ecx
-; AVX512BW-NEXT:    vpextrb $0, %xmm2, %eax
-; AVX512BW-NEXT:    vpextrb $0, %xmm3, %edx
-; AVX512BW-NEXT:    mulb %dl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vmovd %eax, %xmm4
-; AVX512BW-NEXT:    vpinsrb $1, %ecx, %xmm4, %xmm4
-; AVX512BW-NEXT:    vpextrb $2, %xmm2, %eax
-; AVX512BW-NEXT:    vpextrb $2, %xmm3, %ecx
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $2, %eax, %xmm4, %xmm4
-; AVX512BW-NEXT:    vpextrb $3, %xmm2, %eax
-; AVX512BW-NEXT:    vpextrb $3, %xmm3, %ecx
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $3, %eax, %xmm4, %xmm4
-; AVX512BW-NEXT:    vpextrb $4, %xmm2, %eax
-; AVX512BW-NEXT:    vpextrb $4, %xmm3, %ecx
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $4, %eax, %xmm4, %xmm4
-; AVX512BW-NEXT:    vpextrb $5, %xmm2, %eax
-; AVX512BW-NEXT:    vpextrb $5, %xmm3, %ecx
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $5, %eax, %xmm4, %xmm4
-; AVX512BW-NEXT:    vpextrb $6, %xmm2, %eax
-; AVX512BW-NEXT:    vpextrb $6, %xmm3, %ecx
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $6, %eax, %xmm4, %xmm4
-; AVX512BW-NEXT:    vpextrb $7, %xmm2, %eax
-; AVX512BW-NEXT:    vpextrb $7, %xmm3, %ecx
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $7, %eax, %xmm4, %xmm4
-; AVX512BW-NEXT:    vpextrb $8, %xmm2, %eax
-; AVX512BW-NEXT:    vpextrb $8, %xmm3, %ecx
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $8, %eax, %xmm4, %xmm4
-; AVX512BW-NEXT:    vpextrb $9, %xmm2, %eax
-; AVX512BW-NEXT:    vpextrb $9, %xmm3, %ecx
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $9, %eax, %xmm4, %xmm4
-; AVX512BW-NEXT:    vpextrb $10, %xmm2, %eax
-; AVX512BW-NEXT:    vpextrb $10, %xmm3, %ecx
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $10, %eax, %xmm4, %xmm4
-; AVX512BW-NEXT:    vpextrb $11, %xmm2, %eax
-; AVX512BW-NEXT:    vpextrb $11, %xmm3, %ecx
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $11, %eax, %xmm4, %xmm4
-; AVX512BW-NEXT:    vpextrb $12, %xmm2, %eax
-; AVX512BW-NEXT:    vpextrb $12, %xmm3, %ecx
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $12, %eax, %xmm4, %xmm4
-; AVX512BW-NEXT:    vpextrb $13, %xmm2, %eax
-; AVX512BW-NEXT:    vpextrb $13, %xmm3, %ecx
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $13, %eax, %xmm4, %xmm4
-; AVX512BW-NEXT:    vpextrb $14, %xmm2, %eax
-; AVX512BW-NEXT:    vpextrb $14, %xmm3, %ecx
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $14, %eax, %xmm4, %xmm4
-; AVX512BW-NEXT:    vpextrb $15, %xmm2, %eax
-; AVX512BW-NEXT:    vpextrb $15, %xmm3, %ecx
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $15, %eax, %xmm4, %xmm2
-; AVX512BW-NEXT:    vextracti32x4 $2, %zmm0, %xmm3
-; AVX512BW-NEXT:    vpextrb $1, %xmm3, %eax
-; AVX512BW-NEXT:    vextracti32x4 $2, %zmm1, %xmm4
-; AVX512BW-NEXT:    vpextrb $1, %xmm4, %ecx
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %ecx
-; AVX512BW-NEXT:    vpextrb $0, %xmm3, %eax
-; AVX512BW-NEXT:    vpextrb $0, %xmm4, %edx
-; AVX512BW-NEXT:    mulb %dl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vmovd %eax, %xmm5
-; AVX512BW-NEXT:    vpinsrb $1, %ecx, %xmm5, %xmm5
-; AVX512BW-NEXT:    vpextrb $2, %xmm3, %eax
-; AVX512BW-NEXT:    vpextrb $2, %xmm4, %ecx
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $2, %eax, %xmm5, %xmm5
-; AVX512BW-NEXT:    vpextrb $3, %xmm3, %eax
-; AVX512BW-NEXT:    vpextrb $3, %xmm4, %ecx
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $3, %eax, %xmm5, %xmm5
-; AVX512BW-NEXT:    vpextrb $4, %xmm3, %eax
-; AVX512BW-NEXT:    vpextrb $4, %xmm4, %ecx
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $4, %eax, %xmm5, %xmm5
-; AVX512BW-NEXT:    vpextrb $5, %xmm3, %eax
-; AVX512BW-NEXT:    vpextrb $5, %xmm4, %ecx
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $5, %eax, %xmm5, %xmm5
-; AVX512BW-NEXT:    vpextrb $6, %xmm3, %eax
-; AVX512BW-NEXT:    vpextrb $6, %xmm4, %ecx
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $6, %eax, %xmm5, %xmm5
-; AVX512BW-NEXT:    vpextrb $7, %xmm3, %eax
-; AVX512BW-NEXT:    vpextrb $7, %xmm4, %ecx
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $7, %eax, %xmm5, %xmm5
-; AVX512BW-NEXT:    vpextrb $8, %xmm3, %eax
-; AVX512BW-NEXT:    vpextrb $8, %xmm4, %ecx
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $8, %eax, %xmm5, %xmm5
-; AVX512BW-NEXT:    vpextrb $9, %xmm3, %eax
-; AVX512BW-NEXT:    vpextrb $9, %xmm4, %ecx
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $9, %eax, %xmm5, %xmm5
-; AVX512BW-NEXT:    vpextrb $10, %xmm3, %eax
-; AVX512BW-NEXT:    vpextrb $10, %xmm4, %ecx
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $10, %eax, %xmm5, %xmm5
-; AVX512BW-NEXT:    vpextrb $11, %xmm3, %eax
-; AVX512BW-NEXT:    vpextrb $11, %xmm4, %ecx
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $11, %eax, %xmm5, %xmm5
-; AVX512BW-NEXT:    vpextrb $12, %xmm3, %eax
-; AVX512BW-NEXT:    vpextrb $12, %xmm4, %ecx
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $12, %eax, %xmm5, %xmm5
-; AVX512BW-NEXT:    vpextrb $13, %xmm3, %eax
-; AVX512BW-NEXT:    vpextrb $13, %xmm4, %ecx
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $13, %eax, %xmm5, %xmm5
-; AVX512BW-NEXT:    vpextrb $14, %xmm3, %eax
-; AVX512BW-NEXT:    vpextrb $14, %xmm4, %ecx
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $14, %eax, %xmm5, %xmm5
-; AVX512BW-NEXT:    vpextrb $15, %xmm3, %eax
-; AVX512BW-NEXT:    vpextrb $15, %xmm4, %ecx
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $15, %eax, %xmm5, %xmm3
-; AVX512BW-NEXT:    vinserti128 $1, %xmm2, %ymm3, %ymm2
-; AVX512BW-NEXT:    vextracti32x4 $1, %zmm0, %xmm3
-; AVX512BW-NEXT:    vpextrb $1, %xmm3, %eax
-; AVX512BW-NEXT:    vextracti32x4 $1, %zmm1, %xmm4
-; AVX512BW-NEXT:    vpextrb $1, %xmm4, %ecx
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %ecx
-; AVX512BW-NEXT:    vpextrb $0, %xmm3, %eax
-; AVX512BW-NEXT:    vpextrb $0, %xmm4, %edx
-; AVX512BW-NEXT:    mulb %dl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vmovd %eax, %xmm5
-; AVX512BW-NEXT:    vpinsrb $1, %ecx, %xmm5, %xmm5
-; AVX512BW-NEXT:    vpextrb $2, %xmm3, %eax
-; AVX512BW-NEXT:    vpextrb $2, %xmm4, %ecx
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $2, %eax, %xmm5, %xmm5
-; AVX512BW-NEXT:    vpextrb $3, %xmm3, %eax
-; AVX512BW-NEXT:    vpextrb $3, %xmm4, %ecx
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $3, %eax, %xmm5, %xmm5
-; AVX512BW-NEXT:    vpextrb $4, %xmm3, %eax
-; AVX512BW-NEXT:    vpextrb $4, %xmm4, %ecx
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $4, %eax, %xmm5, %xmm5
-; AVX512BW-NEXT:    vpextrb $5, %xmm3, %eax
-; AVX512BW-NEXT:    vpextrb $5, %xmm4, %ecx
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $5, %eax, %xmm5, %xmm5
-; AVX512BW-NEXT:    vpextrb $6, %xmm3, %eax
-; AVX512BW-NEXT:    vpextrb $6, %xmm4, %ecx
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $6, %eax, %xmm5, %xmm5
-; AVX512BW-NEXT:    vpextrb $7, %xmm3, %eax
-; AVX512BW-NEXT:    vpextrb $7, %xmm4, %ecx
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $7, %eax, %xmm5, %xmm5
-; AVX512BW-NEXT:    vpextrb $8, %xmm3, %eax
-; AVX512BW-NEXT:    vpextrb $8, %xmm4, %ecx
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $8, %eax, %xmm5, %xmm5
-; AVX512BW-NEXT:    vpextrb $9, %xmm3, %eax
-; AVX512BW-NEXT:    vpextrb $9, %xmm4, %ecx
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $9, %eax, %xmm5, %xmm5
-; AVX512BW-NEXT:    vpextrb $10, %xmm3, %eax
-; AVX512BW-NEXT:    vpextrb $10, %xmm4, %ecx
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $10, %eax, %xmm5, %xmm5
-; AVX512BW-NEXT:    vpextrb $11, %xmm3, %eax
-; AVX512BW-NEXT:    vpextrb $11, %xmm4, %ecx
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $11, %eax, %xmm5, %xmm5
-; AVX512BW-NEXT:    vpextrb $12, %xmm3, %eax
-; AVX512BW-NEXT:    vpextrb $12, %xmm4, %ecx
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $12, %eax, %xmm5, %xmm5
-; AVX512BW-NEXT:    vpextrb $13, %xmm3, %eax
-; AVX512BW-NEXT:    vpextrb $13, %xmm4, %ecx
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $13, %eax, %xmm5, %xmm5
-; AVX512BW-NEXT:    vpextrb $14, %xmm3, %eax
-; AVX512BW-NEXT:    vpextrb $14, %xmm4, %ecx
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $14, %eax, %xmm5, %xmm5
-; AVX512BW-NEXT:    vpextrb $15, %xmm3, %eax
-; AVX512BW-NEXT:    vpextrb $15, %xmm4, %ecx
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $15, %eax, %xmm5, %xmm3
-; AVX512BW-NEXT:    vpextrb $1, %xmm0, %eax
-; AVX512BW-NEXT:    vpextrb $1, %xmm1, %ecx
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %ecx
-; AVX512BW-NEXT:    vpextrb $0, %xmm0, %eax
-; AVX512BW-NEXT:    vpextrb $0, %xmm1, %edx
-; AVX512BW-NEXT:    mulb %dl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vmovd %eax, %xmm4
-; AVX512BW-NEXT:    vpinsrb $1, %ecx, %xmm4, %xmm4
-; AVX512BW-NEXT:    vpextrb $2, %xmm0, %eax
-; AVX512BW-NEXT:    vpextrb $2, %xmm1, %ecx
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $2, %eax, %xmm4, %xmm4
-; AVX512BW-NEXT:    vpextrb $3, %xmm0, %eax
-; AVX512BW-NEXT:    vpextrb $3, %xmm1, %ecx
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $3, %eax, %xmm4, %xmm4
-; AVX512BW-NEXT:    vpextrb $4, %xmm0, %eax
-; AVX512BW-NEXT:    vpextrb $4, %xmm1, %ecx
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $4, %eax, %xmm4, %xmm4
-; AVX512BW-NEXT:    vpextrb $5, %xmm0, %eax
-; AVX512BW-NEXT:    vpextrb $5, %xmm1, %ecx
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $5, %eax, %xmm4, %xmm4
-; AVX512BW-NEXT:    vpextrb $6, %xmm0, %eax
-; AVX512BW-NEXT:    vpextrb $6, %xmm1, %ecx
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $6, %eax, %xmm4, %xmm4
-; AVX512BW-NEXT:    vpextrb $7, %xmm0, %eax
-; AVX512BW-NEXT:    vpextrb $7, %xmm1, %ecx
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $7, %eax, %xmm4, %xmm4
-; AVX512BW-NEXT:    vpextrb $8, %xmm0, %eax
-; AVX512BW-NEXT:    vpextrb $8, %xmm1, %ecx
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $8, %eax, %xmm4, %xmm4
-; AVX512BW-NEXT:    vpextrb $9, %xmm0, %eax
-; AVX512BW-NEXT:    vpextrb $9, %xmm1, %ecx
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $9, %eax, %xmm4, %xmm4
-; AVX512BW-NEXT:    vpextrb $10, %xmm0, %eax
-; AVX512BW-NEXT:    vpextrb $10, %xmm1, %ecx
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $10, %eax, %xmm4, %xmm4
-; AVX512BW-NEXT:    vpextrb $11, %xmm0, %eax
-; AVX512BW-NEXT:    vpextrb $11, %xmm1, %ecx
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $11, %eax, %xmm4, %xmm4
-; AVX512BW-NEXT:    vpextrb $12, %xmm0, %eax
-; AVX512BW-NEXT:    vpextrb $12, %xmm1, %ecx
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $12, %eax, %xmm4, %xmm4
-; AVX512BW-NEXT:    vpextrb $13, %xmm0, %eax
-; AVX512BW-NEXT:    vpextrb $13, %xmm1, %ecx
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $13, %eax, %xmm4, %xmm4
-; AVX512BW-NEXT:    vpextrb $14, %xmm0, %eax
-; AVX512BW-NEXT:    vpextrb $14, %xmm1, %ecx
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $14, %eax, %xmm4, %xmm4
-; AVX512BW-NEXT:    vpextrb $15, %xmm0, %eax
-; AVX512BW-NEXT:    vpextrb $15, %xmm1, %ecx
-; AVX512BW-NEXT:    mulb %cl
-; AVX512BW-NEXT:    movzbl %al, %eax
-; AVX512BW-NEXT:    vpinsrb $15, %eax, %xmm4, %xmm0
-; AVX512BW-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm0
-; AVX512BW-NEXT:    vinserti64x4 $1, %ymm2, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpmovsxbw %ymm1, %zmm2
+; AVX512BW-NEXT:    vpmovsxbw %ymm0, %zmm3
+; AVX512BW-NEXT:    vpmullw %zmm2, %zmm3, %zmm2
+; AVX512BW-NEXT:    vpmovwb %zmm2, %ymm2
+; AVX512BW-NEXT:    vextracti64x4 $1, %zmm1, %ymm1
+; AVX512BW-NEXT:    vpmovsxbw %ymm1, %zmm1
+; AVX512BW-NEXT:    vextracti64x4 $1, %zmm0, %ymm0
+; AVX512BW-NEXT:    vpmovsxbw %ymm0, %zmm0
+; AVX512BW-NEXT:    vpmullw %zmm1, %zmm0, %zmm0
+; AVX512BW-NEXT:    vpmovwb %zmm0, %ymm0
+; AVX512BW-NEXT:    vinserti64x4 $1, %ymm0, %zmm2, %zmm0
 ; AVX512BW-NEXT:    retq
 entry:
   %A = mul <64 x i8> %i, %j