[llvm] r291670 - X86 CodeGen: Optimized pattern for truncate with unsigned saturation.

Michael Kuperstein via llvm-commits llvm-commits at lists.llvm.org
Tue Jan 24 13:23:40 PST 2017


The branch name is release_40, AFAIK.
As to revert vs. fix - I defer to Hans / Craig.

Michael

On Tue, Jan 24, 2017 at 11:29 AM, Demikhovsky, Elena <
elena.demikhovsky at intel.com> wrote:

> I propose to fix, it will be easier than revert. But it’s not so
> important, Hans’s decision.
>
> I’ll need to know the branch name in order to make changes.
>
>
>
> -          * Elena*
>
>
>
> *From:* Michael Kuperstein [mailto:mkuper at google.com]
> *Sent:* Tuesday, January 24, 2017 21:09
> *To:* Demikhovsky, Elena <elena.demikhovsky at intel.com>; Hans Wennborg <
> hans at chromium.org>
>
> *Cc:* llvm-commits <llvm-commits at lists.llvm.org>
> *Subject:* Re: [llvm] r291670 - X86 CodeGen: Optimized pattern for
> truncate with unsigned saturation.
>
>
>
> I just noticed the original commit is before the 4.0 branch-point, but my
> revert was after.
>
> Do we want to revert r291670 in 4.0? Or merge the fixed version (r292479)?
>
>
>
> Hans, Elena?
>
>
>
> On Thu, Jan 19, 2017 at 10:47 AM, Michael Kuperstein <mkuper at google.com>
> wrote:
>
> Hi Elena,
>
>
>
> Thanks for the fix.
>
>
>
> Regarding the revert - in this case, we're talking about:
>
>
>
> 1) A recent commit,
>
> 2) that has nothing else layered on top of it (except for whitespace
> changes)
>
> 3) is a performance improvement that causes a correctness regression,
>
> 4) the crasher is reduced from real code, not a synthetic test-case,
>
> 5) and has a small IR reproducer.
>
>
>
> I really think that in such cases it's worth keeping trunk clean, at the
> cost of the original commiter having to reverse-merge the revert before
> fixing the bug.
>
>
>
> Thanks,
>
>   Michael
>
>
>
> On Thu, Jan 19, 2017 at 4:49 AM, Demikhovsky, Elena <
> elena.demikhovsky at intel.com> wrote:
>
> Fixed and recommitted in r292479.
>
>
>
> I’d prefer that you’ll not revert the failing commit, but wait for a few
> days. It will be easier for me to fix.
>
> (If it is not a buildbot failure, of course. But these failures I can see
> myself)
>
> We also find regressions in our internal testing from time to time,
> PR31671, for example. We submit a PR, notify the owner, and let him to fix
> the bug.
>
>
>
> Thanks.
>
>
>
> -          * Elena*
>
>
>
> *From:* Michael Kuperstein [mailto:mkuper at google.com]
> *Sent:* Thursday, January 19, 2017 01:19
> *To:* Demikhovsky, Elena <elena.demikhovsky at intel.com>
> *Cc:* llvm-commits <llvm-commits at lists.llvm.org>
> *Subject:* Re: [llvm] r291670 - X86 CodeGen: Optimized pattern for
> truncate with unsigned saturation.
>
>
>
> Hi Elena,
>
>
>
> This still crashes in more complex cases. I've reverted in r292444, see
> PR31589 for the reproducer.
>
>
>
> Thanks,
>
>   Michael
>
>
>
> On Wed, Jan 11, 2017 at 4:59 AM, Elena Demikhovsky via llvm-commits <
> llvm-commits at lists.llvm.org> wrote:
>
> Author: delena
> Date: Wed Jan 11 06:59:32 2017
> New Revision: 291670
>
> URL: http://llvm.org/viewvc/llvm-project?rev=291670&view=rev
> Log:
> X86 CodeGen: Optimized pattern for truncate with unsigned saturation.
>
> DAG patterns optimization: truncate + unsigned saturation supported by
> VPMOVUS* instructions in AVX-512.
> And VPACKUS* instructions on SEE* targets.
>
> Differential Revision: https://reviews.llvm.org/D28216
>
>
> Modified:
>     llvm/trunk/lib/Target/X86/X86ISelLowering.cpp
>     llvm/trunk/test/CodeGen/X86/avx-trunc.ll
>     llvm/trunk/test/CodeGen/X86/avx512-trunc.ll
>
> Modified: llvm/trunk/lib/Target/X86/X86ISelLowering.cpp
> URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/
> X86/X86ISelLowering.cpp?rev=291670&r1=291669&r2=291670&view=diff
> ============================================================
> ==================
> --- llvm/trunk/lib/Target/X86/X86ISelLowering.cpp (original)
> +++ llvm/trunk/lib/Target/X86/X86ISelLowering.cpp Wed Jan 11 06:59:32 2017
> @@ -31220,6 +31220,93 @@ static SDValue foldVectorXorShiftIntoCmp
>    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");
> +
> +  APInt C;
> +  if (ISD::isConstantSplatVector(In.getOperand(1).getNode(), C)) {
> +    // 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) ?
> In.getOperand(0) :
> +      SDValue();
> +  }
> +  return 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.
> @@ -31786,6 +31873,12 @@ static SDValue combineStore(SDNode *N, S
>                            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");
> @@ -32406,6 +32499,10 @@ static SDValue combineTruncate(SDNode *N
>    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) {
>
> Modified: llvm/trunk/test/CodeGen/X86/avx-trunc.ll
> URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/
> CodeGen/X86/avx-trunc.ll?rev=291670&r1=291669&r2=291670&view=diff
> ============================================================
> ==================
> --- llvm/trunk/test/CodeGen/X86/avx-trunc.ll (original)
> +++ llvm/trunk/test/CodeGen/X86/avx-trunc.ll Wed Jan 11 06:59:32 2017
> @@ -39,3 +39,29 @@ define <16 x i8> @trunc_16_8(<16 x i16>
>    %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
> +}
>
> Modified: llvm/trunk/test/CodeGen/X86/avx512-trunc.ll
> URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/
> CodeGen/X86/avx512-trunc.ll?rev=291670&r1=291669&r2=291670&view=diff
> ============================================================
> ==================
> --- llvm/trunk/test/CodeGen/X86/avx512-trunc.ll (original)
> +++ llvm/trunk/test/CodeGen/X86/avx512-trunc.ll Wed Jan 11 06:59:32 2017
> @@ -500,3 +500,208 @@ define void @trunc_wb_128_mem(<8 x i16>
>      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 void @usat_trunc_db_1024_mem(<32 x i32> %i, <32 x i8>* %p) {
> +; KNL-LABEL: usat_trunc_db_1024_mem:
> +; KNL:       ## BB#0:
> +; KNL-NEXT:    vpmovusdb %zmm0, %xmm0
> +; KNL-NEXT:    vpmovusdb %zmm1, %xmm1
> +; KNL-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm0
> +; KNL-NEXT:    vmovdqu %ymm0, (%rdi)
> +; KNL-NEXT:    retq
> +;
> +; SKX-LABEL: usat_trunc_db_1024_mem:
> +; 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, (%rdi)
> +; 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>
> +  store <32 x i8>%x6, <32 x i8>* %p, align 1
> +  ret void
> +}
> +
> +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
> +}
> +
>
>
> _______________________________________________
> llvm-commits mailing list
> llvm-commits at lists.llvm.org
> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-commits
>
>
>
> ---------------------------------------------------------------------
> Intel Israel (74) Limited
>
> This e-mail and any attachments may contain confidential material for
> the sole use of the intended recipient(s). Any review or distribution
> by others is strictly prohibited. If you are not the intended
> recipient, please contact the sender and delete all copies.
>
>
>
>
>
> ---------------------------------------------------------------------
> Intel Israel (74) Limited
>
> This e-mail and any attachments may contain confidential material for
> the sole use of the intended recipient(s). Any review or distribution
> by others is strictly prohibited. If you are not the intended
> recipient, please contact the sender and delete all copies.
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20170124/df3d9f22/attachment.html>


More information about the llvm-commits mailing list