AVX-512: Truncate with unsigned saturation.

[delena]

Summary

DAG patterns optimization: truncate + unsigned saturation supported by VPMOVUS* instructions in AVX-512.

[zvi]

Can you please add a test for VPMOVUSQW?

[RKSimon]

Any chance that we can use this to lower to PACKSS/PACKUS on older targets?

[delena]

In D28216#634306, @RKSimon wrote:

Any chance that we can use this to lower to PACKSS/PACKUS on older targets?

yes, of course. I can do this in one of the next patches.

[delena]

Added QW test as Zvi asked.

[zvi]

Thanks, Elena. LGTM.

[delena]

The previous patch was reverted due to a failure (https://llvm.org/bugs/show_bug.cgi?id=31589).

1. I fixed the bug and added a lot of test cases.
2. I promised to Simon to cover non-avx512 cases with VPACKUS. Now it is in. Additional tests are in avx-trunc.ll.

[RKSimon]

In D28216#641199, @delena wrote:

2. I promised to Simon to cover non-avx512 cases with VPACKUS. Now it is in. Additional tests are in avx-trunc.ll.

Thank you!

../lib/Target/X86/X86ISelLowering.cpp

31130

In.getScalarValueSizeInBits()

31144

Can you use APIntOps::isMask here?

APIntOps::isMask(VT.getScalarSizeInBits(), C)

[delena]

../lib/Target/X86/X86ISelLowering.cpp

31144

Yes! I did not see this interface before. Thank you.

[delena]

Some fixes after Simon's comments.

[RKSimon]

LGTM with minors

../lib/Target/X86/X86ISelLowering.cpp

31137

We canonicalize constants to RHS for UMIN/UMAX/SMIN/SMAX (see DAGCombiner::visitIMINMAX) so we only need this case.

../test/CodeGen/X86/avx512-trunc.ll

630

Is it worth adding PR31589 as usat_trunc_db_1024_mem as well?

../lib/Target/X86/X86ISelLowering.cpp

static SDValue foldVectorXorShiftIntoCmp(SDNode *N, SelectionDAG &DAG,
   if (!ShiftAmt || ShiftAmt->getZExtValue() != ShiftEltTy.getSizeInBits() - 1)
     return SDValue();
 
   // Create a greater-than comparison against -1. We don't use the more obvious
   // greater-than-or-equal-to-zero because SSE/AVX don't have that instruction.
   return DAG.getNode(X86ISD::PCMPGT, SDLoc(N), VT, Shift.getOperand(0), Ones);
 }
 
+/// Check if truncation with saturation form type \p SrcVT to \p DstVT 
+/// is valid for the given \p Subtarget.
+static bool isSATValidOnAVX512Subtarget(EVT SrcVT, EVT DstVT,
+                                        const X86Subtarget &Subtarget) {
+  if (!Subtarget.hasAVX512())
+    return false;
+
+  // FIXME: Scalar type may be supported if we move it to vector register.
+  if (!SrcVT.isVector() || !SrcVT.isSimple() || SrcVT.getSizeInBits() > 512)
+    return false;
+
+  EVT SrcElVT = SrcVT.getScalarType();
+  EVT DstElVT = DstVT.getScalarType();
+  if (SrcElVT.getSizeInBits() < 16 || SrcElVT.getSizeInBits() > 64)
+    return false;
+  if (DstElVT.getSizeInBits() < 8 || DstElVT.getSizeInBits() > 32)
+    return false;
+  if (SrcVT.is512BitVector() || Subtarget.hasVLX())
+    return SrcElVT.getSizeInBits() >= 32 || Subtarget.hasBWI();
+  return false;
+}
+
+/// Return true if VPACK* instruction can be used for the given types
+/// and it is avalable on \p Subtarget.
+static bool
+isSATValidOnSSESubtarget(EVT SrcVT, EVT DstVT, const X86Subtarget &Subtarget) {
+  if (Subtarget.hasSSE2())
+    // v16i16 -> v16i8
+    if (SrcVT == MVT::v16i16 && DstVT == MVT::v16i8)
+      return true;
+  if (Subtarget.hasSSE41())
+    // v8i32 -> v8i16
+    if (SrcVT == MVT::v8i32 && DstVT == MVT::v8i16)
+      return true;
+  return false;
+}
+
+/// Detect a pattern of truncation with saturation:
+/// (truncate (umin (x, unsigned_max_of_dest_type)) to dest_type).
+/// Return the source value to be truncated or SDValue() if the pattern was not
+/// matched.
+static SDValue detectUSatPattern(SDValue In, EVT VT) {
+  if (In.getOpcode() != ISD::UMIN)
+    return SDValue();
+
+  //Saturation with truncation. We truncate from InVT to VT.
+  assert(In.getScalarValueSizeInBits() > VT.getScalarSizeInBits() &&
+    "Unexpected types for truncate operation");
+
+  SDValue SrcVal;
+  APInt C;
+  if (ISD::isConstantSplatVector(In.getOperand(0).getNode(), C))
+    SrcVal = In.getOperand(1);
+  else if (ISD::isConstantSplatVector(In.getOperand(1).getNode(), C))
+    SrcVal = In.getOperand(0);
+  else
+    return SDValue();
+
+  // C should be equal to UINT32_MAX / UINT16_MAX / UINT8_MAX according
+  // the element size of the destination type.
+  return APIntOps::isMask(VT.getScalarSizeInBits(), C) ? SrcVal : SDValue();
+}
+
+/// Detect a pattern of truncation with saturation:
+/// (truncate (umin (x, unsigned_max_of_dest_type)) to dest_type).
+/// The types should allow to use VPMOVUS* instruction on AVX512.
+/// Return the source value to be truncated or SDValue() if the pattern was not
+/// matched.
+static SDValue detectAVX512USatPattern(SDValue In, EVT VT,
+                                       const X86Subtarget &Subtarget) {
+  if (!isSATValidOnAVX512Subtarget(In.getValueType(), VT, Subtarget))
+    return SDValue();
+  return detectUSatPattern(In, VT);
+}
+
+static SDValue
+combineTruncateWithUSat(SDValue In, EVT VT, SDLoc &DL, SelectionDAG &DAG,
+                        const X86Subtarget &Subtarget) {
+  SDValue USatVal = detectUSatPattern(In, VT);
+  if (USatVal) {
+    if (isSATValidOnAVX512Subtarget(In.getValueType(), VT, Subtarget))
+      return DAG.getNode(X86ISD::VTRUNCUS, DL, VT, USatVal);
+    if (isSATValidOnSSESubtarget(In.getValueType(), VT, Subtarget)) {
+      SDValue Lo, Hi;
+      std::tie(Lo, Hi) = DAG.SplitVector(USatVal, DL);
+      return DAG.getNode(X86ISD::PACKUS, DL, VT, Lo, Hi);
+    }
+  }
+  return SDValue();
+}
+
 /// This function detects the AVG pattern between vectors of unsigned i8/i16,
 /// which is c = (a + b + 1) / 2, and replace this operation with the efficient
 /// X86ISD::AVG instruction.
 static SDValue detectAVGPattern(SDValue In, EVT VT, SelectionDAG &DAG,
                                 const X86Subtarget &Subtarget,
                                 const SDLoc &DL) {
   if (!VT.isVector() || !VT.isSimple())
     return SDValue();
if (St->isTruncatingStore() && VT.isVector()) {
     // replace the trunc store by a normal store with the result of X86ISD::AVG
     // instruction.
     if (SDValue Avg = detectAVGPattern(St->getValue(), St->getMemoryVT(), DAG,
                                        Subtarget, dl))
       return DAG.getStore(St->getChain(), dl, Avg, St->getBasePtr(),
                           St->getPointerInfo(), St->getAlignment(),
                           St->getMemOperand()->getFlags());
 
+    if (SDValue Val =
+        detectAVX512USatPattern(St->getValue(), St->getMemoryVT(), Subtarget))
+      return EmitTruncSStore(false /* Unsigned saturation */, St->getChain(),
+                             dl, Val, St->getBasePtr(),
+                             St->getMemoryVT(), St->getMemOperand(), DAG);
+
     const TargetLowering &TLI = DAG.getTargetLoweringInfo();
     unsigned NumElems = VT.getVectorNumElements();
     assert(StVT != VT && "Cannot truncate to the same type");
     unsigned FromSz = VT.getScalarSizeInBits();
     unsigned ToSz = StVT.getScalarSizeInBits();
 
     // The truncating store is legal in some cases. For example
     // vpmovqb, vpmovqw, vpmovqd, vpmovdb, vpmovdw
static SDValue combineTruncate(SDNode *N, SelectionDAG &DAG,
   // Attempt to pre-truncate inputs to arithmetic ops instead.
   if (SDValue V = combineTruncatedArithmetic(N, DAG, Subtarget, DL))
     return V;
 
   // Try to detect AVG pattern first.
   if (SDValue Avg = detectAVGPattern(Src, VT, DAG, Subtarget, DL))
     return Avg;
 
+  // Try to combine truncation with unsigned saturation.
+  if (SDValue Val = combineTruncateWithUSat(Src, VT, DL, DAG, Subtarget))
+    return Val;
+
   // The bitcast source is a direct mmx result.
   // Detect bitcasts between i32 to x86mmx
   if (Src.getOpcode() == ISD::BITCAST && VT == MVT::i32) {
     SDValue BCSrc = Src.getOperand(0);
     if (BCSrc.getValueType() == MVT::x86mmx)
       return DAG.getNode(X86ISD::MMX_MOVD2W, DL, MVT::i32, BCSrc);
   }
 

../test/CodeGen/X86/avx-trunc.ll

 ; CHECK-NEXT:    vpshufb %xmm2, %xmm1, %xmm1
 ; CHECK-NEXT:    vpshufb %xmm2, %xmm0, %xmm0
 ; CHECK-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
 ; CHECK-NEXT:    vzeroupper
 ; CHECK-NEXT:    retq
   %B = trunc <16 x i16> %A to <16 x i8>
   ret <16 x i8> %B
 }
+
+define <16 x i8> @usat_trunc_wb_256(<16 x i16> %i) {
+; CHECK-LABEL: usat_trunc_wb_256:
+; CHECK:       # BB#0:
+; CHECK-NEXT:    vextractf128 $1, %ymm0, %xmm1
+; CHECK-NEXT:    vpackuswb %xmm1, %xmm0, %xmm0
+; CHECK-NEXT:    vzeroupper
+; CHECK-NEXT:    retq
+  %x3 = icmp ult <16 x i16> %i, <i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255>
+  %x5 = select <16 x i1> %x3, <16 x i16> %i, <16 x i16> <i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255>
+  %x6 = trunc <16 x i16> %x5 to <16 x i8>
+  ret <16 x i8> %x6
+}
+
+define <8 x i16> @usat_trunc_dw_256(<8 x i32> %i) {
+; CHECK-LABEL: usat_trunc_dw_256:
+; CHECK:       # BB#0:
+; CHECK-NEXT:    vextractf128 $1, %ymm0, %xmm1
+; CHECK-NEXT:    vpackusdw %xmm1, %xmm0, %xmm0
+; CHECK-NEXT:    vzeroupper
+; CHECK-NEXT:    retq
+  %x3 = icmp ult <8 x i32> %i, <i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535>
+  %x5 = select <8 x i1> %x3, <8 x i32> %i, <8 x i32> <i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535>
+  %x6 = trunc <8 x i32> %x5 to <8 x i16>
+  ret <8 x i16> %x6
+}

../test/CodeGen/X86/avx512-trunc.ll

 ; SKX-LABEL: trunc_wb_128_mem:
 ; SKX:       ## BB#0:
 ; SKX-NEXT:    vpmovwb %xmm0, (%rdi)
 ; SKX-NEXT:    retq
     %x = trunc <8 x i16> %i to <8 x i8>
     store <8 x i8> %x, <8 x i8>* %res
     ret void
 }
+
+
+define void @usat_trunc_wb_256_mem(<16 x i16> %i, <16 x i8>* %res) {
+; KNL-LABEL: usat_trunc_wb_256_mem:
+; KNL:       ## BB#0:
+; KNL-NEXT:    vextracti128 $1, %ymm0, %xmm1
+; KNL-NEXT:    vpackuswb %xmm1, %xmm0, %xmm0
+; KNL-NEXT:    vmovdqu %xmm0, (%rdi)
+; KNL-NEXT:    retq
+;
+; SKX-LABEL: usat_trunc_wb_256_mem:
+; SKX:       ## BB#0:
+; SKX-NEXT:    vpmovuswb %ymm0, (%rdi)
+; SKX-NEXT:    retq
+  %x3 = icmp ult <16 x i16> %i, <i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255>
+  %x5 = select <16 x i1> %x3, <16 x i16> %i, <16 x i16> <i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255>
+  %x6 = trunc <16 x i16> %x5 to <16 x i8>
+  store <16 x i8> %x6, <16 x i8>* %res, align 1
+  ret void
+}
+
+define <16 x i8> @usat_trunc_wb_256(<16 x i16> %i) {
+; KNL-LABEL: usat_trunc_wb_256:
+; KNL:       ## BB#0:
+; KNL-NEXT:    vextracti128 $1, %ymm0, %xmm1
+; KNL-NEXT:    vpackuswb %xmm1, %xmm0, %xmm0
+; KNL-NEXT:    retq
+;
+; SKX-LABEL: usat_trunc_wb_256:
+; SKX:       ## BB#0:
+; SKX-NEXT:    vpmovuswb %ymm0, %xmm0
+; SKX-NEXT:    retq
+  %x3 = icmp ult <16 x i16> %i, <i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255>
+  %x5 = select <16 x i1> %x3, <16 x i16> %i, <16 x i16> <i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255>
+  %x6 = trunc <16 x i16> %x5 to <16 x i8>
+  ret <16 x i8> %x6
+}
+
+define void @usat_trunc_wb_128_mem(<8 x i16> %i, <8 x i8>* %res) {
+; KNL-LABEL: usat_trunc_wb_128_mem:
+; KNL:       ## BB#0:
+; KNL-NEXT:    vpminuw {{.*}}(%rip), %xmm0, %xmm0
+; KNL-NEXT:    vpshufb {{.*#+}} xmm0 = xmm0[0,2,4,6,8,10,12,14,u,u,u,u,u,u,u,u]
+; KNL-NEXT:    vmovq %xmm0, (%rdi)
+; KNL-NEXT:    retq
+;
+; SKX-LABEL: usat_trunc_wb_128_mem:
+; SKX:       ## BB#0:
+; SKX-NEXT:    vpmovuswb %xmm0, (%rdi)
+; SKX-NEXT:    retq
+  %x3 = icmp ult <8 x i16> %i, <i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255>
+  %x5 = select <8 x i1> %x3, <8 x i16> %i, <8 x i16> <i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255>
+  %x6 = trunc <8 x i16> %x5 to <8 x i8>
+  store <8 x i8> %x6, <8 x i8>* %res, align 1
+  ret void
+}
+
+define void @usat_trunc_db_512_mem(<16 x i32> %i, <16 x i8>* %res) {
+; ALL-LABEL: usat_trunc_db_512_mem:
+; ALL:       ## BB#0:
+; ALL-NEXT:    vpmovusdb %zmm0, (%rdi)
+; ALL-NEXT:    retq
+  %x3 = icmp ult <16 x i32> %i, <i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255>
+  %x5 = select <16 x i1> %x3, <16 x i32> %i, <16 x i32> <i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255>
+  %x6 = trunc <16 x i32> %x5 to <16 x i8>
+  store <16 x i8> %x6, <16 x i8>* %res, align 1
+  ret void
+}
+
+define void @usat_trunc_qb_512_mem(<8 x i64> %i, <8 x i8>* %res) {
+; ALL-LABEL: usat_trunc_qb_512_mem:
+; ALL:       ## BB#0:
+; ALL-NEXT:    vpmovusqb %zmm0, (%rdi)
+; ALL-NEXT:    retq
+  %x3 = icmp ult <8 x i64> %i, <i64 255, i64 255, i64 255, i64 255, i64 255, i64 255, i64 255, i64 255>
+  %x5 = select <8 x i1> %x3, <8 x i64> %i, <8 x i64> <i64 255, i64 255, i64 255, i64 255, i64 255, i64 255, i64 255, i64 255>
+  %x6 = trunc <8 x i64> %x5 to <8 x i8>
+  store <8 x i8> %x6, <8 x i8>* %res, align 1
+  ret void
+}
+
+define void @usat_trunc_qd_512_mem(<8 x i64> %i, <8 x i32>* %res) {
+; ALL-LABEL: usat_trunc_qd_512_mem:
+; ALL:       ## BB#0:
+; ALL-NEXT:    vpmovusqd %zmm0, (%rdi)
+; ALL-NEXT:    retq
+  %x3 = icmp ult <8 x i64> %i, <i64 4294967295, i64 4294967295, i64 4294967295, i64 4294967295, i64 4294967295, i64 4294967295, i64 4294967295, i64 4294967295>
+  %x5 = select <8 x i1> %x3, <8 x i64> %i, <8 x i64> <i64 4294967295, i64 4294967295, i64 4294967295, i64 4294967295, i64 4294967295, i64 4294967295, i64 4294967295, i64 4294967295>
+  %x6 = trunc <8 x i64> %x5 to <8 x i32>
+  store <8 x i32> %x6, <8 x i32>* %res, align 1
+  ret void
+}
+
+define void @usat_trunc_qw_512_mem(<8 x i64> %i, <8 x i16>* %res) {
+; ALL-LABEL: usat_trunc_qw_512_mem:
+; ALL:       ## BB#0:
+; ALL-NEXT:    vpmovusqw %zmm0, (%rdi)
+; ALL-NEXT:    retq
+  %x3 = icmp ult <8 x i64> %i, <i64 65535, i64 65535, i64 65535, i64 65535, i64 65535, i64 65535, i64 65535, i64 65535>
+  %x5 = select <8 x i1> %x3, <8 x i64> %i, <8 x i64> <i64 65535, i64 65535, i64 65535, i64 65535, i64 65535, i64 65535, i64 65535, i64 65535>
+  %x6 = trunc <8 x i64> %x5 to <8 x i16>
+  store <8 x i16> %x6, <8 x i16>* %res, align 1
+  ret void
+}
+
+define <32 x i8> @usat_trunc_db_1024(<32 x i32> %i) {
+; KNL-LABEL: usat_trunc_db_1024:
+; KNL:       ## BB#0:
+; KNL-NEXT:    vpmovusdb %zmm0, %xmm0
+; KNL-NEXT:    vpmovusdb %zmm1, %xmm1
+; KNL-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm0
+; KNL-NEXT:    retq
+;
+; SKX-LABEL: usat_trunc_db_1024:
+; SKX:       ## BB#0:
+; SKX-NEXT:    vpbroadcastd {{.*}}(%rip), %zmm2
+; SKX-NEXT:    vpminud %zmm2, %zmm1, %zmm1
+; SKX-NEXT:    vpminud %zmm2, %zmm0, %zmm0
+; SKX-NEXT:    vpmovdw %zmm0, %ymm0
+; SKX-NEXT:    vpmovdw %zmm1, %ymm1
+; SKX-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
+; SKX-NEXT:    vpmovwb %zmm0, %ymm0
+; SKX-NEXT:    retq
+  %x3 = icmp ult <32 x i32> %i, <i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255>
+  %x5 = select <32 x i1> %x3, <32 x i32> %i, <32 x i32> <i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255>
+  %x6 = trunc <32 x i32> %x5 to <32 x i8>
+  ret <32 x i8> %x6
+}
+
+define <16 x i16> @usat_trunc_dw_512(<16 x i32> %i) {
+; ALL-LABEL: usat_trunc_dw_512:
+; ALL:       ## BB#0:
+; ALL-NEXT:    vpmovusdw %zmm0, %ymm0
+; ALL-NEXT:    retq
+  %x3 = icmp ult <16 x i32> %i, <i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535>
+  %x5 = select <16 x i1> %x3, <16 x i32> %i, <16 x i32> <i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535>
+  %x6 = trunc <16 x i32> %x5 to <16 x i16>
+  ret <16 x i16> %x6
+}
+
+define <8 x i8> @usat_trunc_wb_128(<8 x i16> %i) {
+; ALL-LABEL: usat_trunc_wb_128:
+; ALL:       ## BB#0:
+; ALL-NEXT:    vpminuw {{.*}}(%rip), %xmm0, %xmm0
+; ALL-NEXT:    retq
+  %x3 = icmp ult <8 x i16> %i, <i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255>
+  %x5 = select <8 x i1> %x3, <8 x i16> %i, <8 x i16> <i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255>
+  %x6 = trunc <8 x i16> %x5 to <8 x i8>
+  ret <8 x i8>%x6
+}
+
+define <16 x i16> @usat_trunc_qw_1024(<16 x i64> %i) {
+; KNL-LABEL: usat_trunc_qw_1024:
+; KNL:       ## BB#0:
+; KNL-NEXT:    vpbroadcastq {{.*}}(%rip), %zmm2
+; KNL-NEXT:    vpminuq %zmm2, %zmm1, %zmm1
+; KNL-NEXT:    vpminuq %zmm2, %zmm0, %zmm0
+; KNL-NEXT:    vpmovqd %zmm0, %ymm0
+; KNL-NEXT:    vpmovqd %zmm1, %ymm1
+; KNL-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
+; KNL-NEXT:    vpmovdw %zmm0, %ymm0
+; KNL-NEXT:    retq
+;
+; SKX-LABEL: usat_trunc_qw_1024:
+; SKX:       ## BB#0:
+; SKX-NEXT:    vpbroadcastq {{.*}}(%rip), %zmm2
+; SKX-NEXT:    vpminuq %zmm2, %zmm1, %zmm1
+; SKX-NEXT:    vpminuq %zmm2, %zmm0, %zmm0
+; SKX-NEXT:    vpmovqd %zmm0, %ymm0
+; SKX-NEXT:    vpmovqd %zmm1, %ymm1
+; SKX-NEXT:    vinserti32x8 $1, %ymm1, %zmm0, %zmm0
+; SKX-NEXT:    vpmovdw %zmm0, %ymm0
+; SKX-NEXT:    retq
+  %x3 = icmp ult <16 x i64> %i, <i64 65535, i64 65535, i64 65535, i64 65535, i64 65535, i64 65535, i64 65535, i64 65535, i64 65535, i64 65535, i64 65535, i64 65535, i64 65535, i64 65535, i64 65535, i64 65535>
+  %x5 = select <16 x i1> %x3, <16 x i64> %i, <16 x i64> <i64 65535, i64 65535, i64 65535, i64 65535, i64 65535, i64 65535, i64 65535, i64 65535, i64 65535, i64 65535, i64 65535, i64 65535, i64 65535, i64 65535, i64 65535, i64 65535>
+  %x6 = trunc <16 x i64> %x5 to <16 x i16>
+  ret <16 x i16> %x6
+}
+